Servo cable bending stress adaptive prediction method based on multi-sensor fusion

Abstract

The invention relates to the technical field of industrial equipment state monitoring, and provides a servo cable bending stress prediction method based on multi-sensor fusion. Aiming at the problems of single data dimension, poor generalization ability of a machine learning model and insufficient real-time performance in the prior art, the method comprises the following steps of: firstly, acquiring and preprocessing a multi-source historical data set by deploying a multi-modal sensor; and then, constructing a neural network fusion layer to fuse the preprocessed data set, and generating comprehensive time-frequency domain fusion data. Thirdly, constructing a trainable deep neural network hybrid architecture, performing deep feature extraction on the fused data, and executing pattern correlation analysis to obtain a comprehensive correlation matrix; and finally, directly inputting new data acquired in real time into the prediction neural network, and performing time sequence prediction reasoning in combination with the comprehensive incidence matrix to obtain final stress prediction data. According to the method, the accuracy of predicting the bending stress of the servo cable is remarkably improved, and the adaptability and robustness of the system are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial equipment condition monitoring, and particularly relates to an adaptive prediction method for the bending stress of servo cables based on multi-sensor fusion. Background Art

[0002] As a key component of industrial automation equipment, servo cables are prone to fatigue, aging, and even fracture under frequent complex mechanical stresses such as bending and torsion, resulting in production interruptions, economic losses, and safety hazards. Traditional passive maintenance or periodic replacement strategies are costly and cannot effectively avoid sudden failures, making it difficult to meet the requirements of modern industry for high reliability and predictive maintenance.

[0003] In the prior art, some methods use single or limited sensors combined with shallow neural networks for stress prediction. Although such methods are easy to deploy, under complex working conditions, due to the failure to fully consider the comprehensive influence of multi-physical field coupling effects such as temperature, load current, and vibration on cable stress, their prediction accuracy is insufficient. At the same time, the shallow neural network architecture is difficult to comprehensively capture these complex non-linear correlations and dynamic changes, limiting its application effect under variable working conditions.

[0004] To improve performance, some research has turned to more complex traditional machine learning algorithms and combined optimization means such as gradient descent. These methods can theoretically optimize model parameters, but they usually rely on time-consuming offline training and are difficult to meet the high requirements of industrial scenarios for real-time monitoring, millisecond-level response, and model dynamic update. Especially when facing sudden changes in working conditions, their prediction ability is limited.

[0005] In recent years, deep learning neural network technologies, especially deep temporal convolutional neural networks and bidirectional long short-term memory neural networks, have shown great potential in processing temporal data and extracting complex features, significantly improving the prediction accuracy of cable bending stress and being able to better capture non-linear dynamics and long-range dependencies. Nevertheless, there is still room for improvement in the current application in terms of feature fusion efficiency. Traditional fusion methods fail to fully explore the deep spatio-temporal correlations between multi-source sensor data and their actual contribution to the prediction target, which may lead to information redundancy or loss of key features. In addition, deep neural networks are often regarded as "black boxes" because their internal decision-making mechanisms are not transparent and it is difficult to explain the specific impact of each input factor on the prediction result, which to a certain extent limits their wide application in industrial fault diagnosis and predictive maintenance scenarios that require high credibility. Summary of the Invention

[0006] The technical problem to be solved by the present invention is the problems of one-sided data, poor model adaptability, and low prediction accuracy in traditional stress prediction technologies.

[0007] To achieve the above object, the present invention provides the following technical solutions: A servo cable bending stress adaptive prediction method based on multi-sensor fusion, including.

[0008] Deploy a multi-modal sensor array in the bending area of the servo cable to obtain the first historical data set, including bending strain data, temperature gradient data, vibration data, load current data and stress data; and perform preprocessing operations of online cleaning and spatio-temporal calibration to obtain the second historical data set; Construct a deep neural network fusion layer of the entropy weight method and the attention mechanism to fuse the second historical data set to obtain comprehensive time-frequency domain fusion data; Construct a trainable deep neural network hybrid architecture based on a deep temporal convolutional neural network and a bidirectional long short-term memory neural network to extract features from the comprehensive time-frequency domain fusion data, generate comprehensive time-frequency domain fusion features; perform correlation analysis on the time-frequency domain fusion features through the neural network architecture to obtain a time-domain feature correlation matrix and a frequency-domain feature correlation matrix, and then obtain a comprehensive correlation matrix; Obtain bending strain data, temperature gradient data, vibration data and load current data in real time, and obtain stress prediction data according to these four data and the comprehensive correlation matrix.

[0009] Preferably, the preprocessing process includes: Perform online cleaning on the obtained first historical data, use the digital filtering method of wavelet threshold denoising to eliminate high-frequency vibration interference, and perform five-layer decomposition using the db4 wavelet basis. The db4 wavelet basis is a commonly used compactly supported orthogonal wavelet basis with good time-frequency localization characteristics, suitable for multi-scale analysis of signals, and can better match the transient characteristics of strain signals, thereby retaining the effective information of the signal; this method uses five-layer wavelet decomposition, and the selection of the decomposition layer is based on the length and characteristics of the signal. Five-layer decomposition can effectively remove high-frequency noise while retaining the main characteristics of the signal. However, the specific decomposition layer needs to be adjusted according to the specific situation of the signal to achieve the best denoising effect.

[0010] This method performs multi-layer wavelet decomposition operations on the first historical data set to match the transient digital characteristics of the strain signal and retain the effective data; among them, the noise threshold is dynamically calculated through parameter optimization according to the length of the digital signal and the standard deviation of the noise. The wavelet coefficients at each scale are obtained by performing wavelet decomposition on the digital signal through a soft threshold function to suppress noise interference. The soft threshold is calculated according to the general threshold rule combined with the estimated value of the standard deviation of the noise at each layer. The specific steps are as follows: Perform j-layer wavelet decomposition on the input signal, and perform 5-layer decomposition using the db4 wavelet basis to obtain the detail coefficients of each decomposition layer j (j = 1,..., J) For the detail coefficients of each decomposition layer j , the median absolute deviation (MAD) method is used to estimate its noise standard deviation , the noise standard deviation The calculation formula is as follows: ; Among them, .

[0011] Combined with the estimated value of its noise standard deviation and the length of the detail coefficients of this layer for calculation. The threshold calculation formula for each decomposition layer is: ; Among them, is the threshold calculated for the j-th decomposition layer, is the estimated value of the noise standard deviation of the detail coefficients of the j-th decomposition layer, is the signal length of the detail coefficients of the j-th decomposition layer.

[0012] Then, the data is subjected to spatio-temporal calibration processing. For data records from sensors with different sampling rates, where the sampling rate of the temperature gradient sensor is 1 Hz and the sampling rate of the current sensor is 50 kHz, the numerical interpolation algorithm is used to align their digital timestamps. Through cubic spline interpolation, it is ensured that the timestamp alignment error is less than or equal to 0.05 milliseconds, and by performing a sliding window normalization operation on the digital data, the window length is set to 10 seconds, the normalization range is from 0 to 1, and a small protection constant is added to avoid division by zero to eliminate the dimensional differences between different data. The normalization calculation process is as follows: for each data point within the window, its normalized value is equal to minus the window minimum value and then divided by the window maximum value minus the window minimum value plus the protection constant : ; Among them, .

[0013] After the strain data and temperature data are normalized to the range of [0, 1], they are converted into dimensionless numerical values, which are comparable during feature fusion; for the high-frequency characteristics of the acceleration sensor, an additional 0.1-second short-time window analysis is performed to extract the local statistical characteristics of transient shock events.

[0014] Preferably, the fusion of the second historical dataset by the constructed deep neural network fusion layer of the entropy weight method and the attention mechanism includes: Calculate the information entropy weights of the data obtained by each multimodal sensor in the second historical dataset through the entropy weight method, bin the data into 20 intervals, and then calculate the probability distribution (where \(i\) is the interval and \(j\) is the sensor modality), and use the information entropy formula to calculate the information entropy of the \(j\)th sensor modality ; ; where, , and \(m\) is the number of bins. This \(k\) value is used to normalize the information entropy to the interval \([0, 1]\).

[0015] The information entropy weight of the \(j\)th sensor modality is calculated by the following formula: ; where, is the total number of sensors, is the information entropy of the \(k\)th sensor modality, and when summing, \(k\) traverses all sensors.

[0016] Taking the time-frequency domain features of each multimodal sensor in the second historical dataset as input, based on the attention mechanism, including a single-layer fully connected neural network with 16 hidden units, using the Softmax activation function, taking the time-domain features of each multimodal sensor, such as mean, variance, and peak-to-peak value, as input, and outputting the attention weights representing the feature importance , which represents the attention weight of the \(j\)th sensor modality, and all form a vector . The final sensor fusion weight of the \(j\)th sensor modality is the weighted combination of the information entropy weight of the \(j\)th sensor modality and the attention weight of the \(j\)th sensor modality, using linear weighted averaging with a weight coefficient of 0.5: ; Preferably, the trainable deep neural network hybrid architecture is: The deep temporal convolutional neural network module consists of three dilated convolutional neural network units. The convolutional kernel size of each layer is fixed at 3×1. Through the dilatation factors that multiply layer by layer, where the layer number d = 1, 2, 4, for the input time series data, covering 32 time steps, corresponding to 3.2 seconds of data at a sampling rate of 10 Hz, convolution operations are performed to achieve coverage of a large receptive field of the time series, thereby capturing digital dependencies across different time scales; the bidirectional long short-term memory neural network module is connected after the output of the deep temporal convolutional neural network module. Its hidden layer nodes are configured with 64, and the time step is 10. Through the gating mechanism (such as input gate, forget gate, output gate) and parameter matrix operations, the influence of historical and future data on the current predicted value is learned from the forward and reverse directions respectively; the output layer of the model passes through a fully connected neural network layer and the non-linear activation function ReLU function to ensure that the output is non-negative to generate the predicted stress-related digital data.

[0017] Preferably, the training process of the model includes: Set the initial learning rate value to 0.005, and perform parameter optimization by decaying the learning rate value by 5% every 10 training epochs, that is: ; Among them, is the new learning rate, which is the learning rate value that the model will use in the next training epoch after one decay adjustment. is the current learning rate, which is the learning rate value that the model is currently using before this decay adjustment, that is, the set initial learning rate of 0.005. 5% is the learning rate decay rate.

[0018] The loss function selects the Huber loss function, and the delta parameter is set to 1.0. This function performs piecewise arithmetic operations on the prediction error e. When the absolute value of the error is less than the delta parameter square error calculation is used, and when it is greater than or equal to the delta parameter linear error calculation is used. This combines the calculation of the square error term and the linear error term to evaluate the model performance: ; Among them, is the delta parameter of the Huber loss function, which is a key hyperparameter of the Huber loss function and defines the threshold point for the switching of the loss function behavior; e is the prediction error, which is the difference between the model prediction value and the true value: ; Among them, is the true stress value, which is the actual observed, sensor-collected, or calibrated true bending stress value of the servo cable. In the training dataset, it is provided to the model as a label; is the model-predicted stress value, which is the value output by the constructed hybrid architecture of a deep neural network based on a deep temporal convolutional neural network and a bidirectional long short-term memory neural network for predicting the bending stress of the servo cable.

[0019] The Adam algorithm is selected as the optimizer, with momentum parameters beta1 = 0.9 and beta2 = 0.999. The model parameters, such as the convolutional kernel weights of the TCN and the gating unit weights of the BiLSTM, are updated through the gradient descent algorithm based on backpropagation. In combination with the early stopping strategy, a comparison logic based on the validation set error is used to prevent the model from overfitting, and finally, the model parameters with the best performance during the training process are saved.

[0020] Preferably, the correlation analysis of the time-frequency domain fusion features through the hybrid architecture of the deep neural network yields a time-domain feature correlation matrix and a frequency-domain feature correlation matrix, and further obtaining the comprehensive correlation matrix includes: The hybrid architecture of the neural network obtains the respective time-domain features, such as mean, variance, peak, kurtosis, and frequency-domain features, such as energy spectral density, main frequency, bandwidth, and frequency-domain centroid, of the bending strain data, temperature gradient data, vibration data, and load current data from the comprehensive time-frequency domain fusion features. Calculate the correlation between the time-domain features and frequency-domain features of each original data and the predicted stress, and generate the time-domain feature correlation matrix and the frequency-domain feature correlation matrix through the Pearson correlation coefficient and the self-attention weights of the neural network. Among them, the Pearson correlation coefficient The calculation formula is: ; Among them, and respectively represent the i-th observation values of two variables, and respectively represent the means of the two variables, and N represents the number of observations.

[0021] The self-attention weights of the neural network: used to capture the complex relationships between features, and the calculation formula is: ; Among them, Q, K, and V respectively represent the query vector, key vector, and value vector, represents the dimension of the key vector, and the softmax function is used to convert the correlation into a probability distribution.

[0022] Finally, the time-domain feature correlation matrix and the frequency-domain feature correlation matrix Perform weighted combination. By performing arithmetic mean or weighted mean on the elements at the corresponding positions, the comprehensive correlation matrix is obtained. , and the calculation formula is: ; where is the weighting coefficient, which can be adjusted according to actual needs. ranges from [0, 1]. When , the comprehensive correlation matrix is completely determined by the frequency-domain feature correlation matrix; when , the comprehensive correlation matrix is completely determined by the time-domain feature correlation matrix; when , the comprehensive correlation matrix is jointly determined by the time-domain feature correlation matrix and the frequency-domain feature correlation matrix.

[0023] Preferably, the method for obtaining stress prediction data based on these four data and the comprehensive correlation matrix is as follows: The bending strain data, temperature gradient data, vibration data, and load current data obtained in real time, after preprocessing and preliminary time-frequency domain feature extraction, obtain the real-time time-frequency domain feature vector , and the real-time time-frequency domain feature vector and the comprehensive correlation matrix , are calculated through the neural network weight mapping layer. Through a linear mapping layer L, whose weights have been learned and derived from during the training stage, map the real-time features to stress feature data: ; Finally, the final stress prediction data is generated through calculation.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Through the deployment of multi-modal sensors and data fusion, the present invention can obtain more comprehensive servo cable status information. Combined with online cleaning and spatio-temporal calibration processing based on the neural network preprocessing layer, the quality and consistency of the input data are significantly improved, laying a solid foundation for subsequent accurate prediction.

[0025] 2. The present invention adopts a dynamic weight allocation strategy that combines the entropy weight method and the attention mechanism, which can adaptively adjust the fusion weights according to the information volume and importance of each sensor data, enhancing the effectiveness of deep neural network feature extraction and the adaptability to working condition changes; at the same time, based on the trainable TCN-BiGRU neural network hybrid architecture, the prediction model can capture long-range dependencies and learn complex propagation laws, thereby improving the prediction accuracy of the servo cable bending stress-related values and the generalization ability of the model.

[0026] 3. Through the correlation analysis of the deep neural network, the present invention can quantify and generate the feature correlation matrices and the comprehensive correlation matrix in the time domain and frequency domain, clearly reveal the specific contributions of different sensor data dimensions to stress prediction, enhance the interpretability of the prediction model, and provide direct weight guidance for real-time stress prediction.

[0027] 4. By using the comprehensive correlation matrix learned by the neural network, the present invention directly guides the stress prediction of real-time multimodal data, realizes the real-time and accuracy of prediction, effectively supports the predictive maintenance of servo cables, and reduces the failure risk and operation and maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0029] Figure 1 It is a flowchart of the state monitoring and prediction of the servo cable of the precision engraving robot proposed in the embodiment of the present invention application; Figure 2 It is a flowchart of the state monitoring and prediction of the servo cable of the single-axis engraving robot proposed in the embodiment of the present invention application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0030] Next, the technical solutions of the present invention will be described in detail and clearly in combination with the specific examples and related drawings of the present invention. It should be clear that the examples described below only represent a part of the present invention, not all of it. Based on these examples, those of ordinary skill in the art can derive other possible embodiments of the present invention without creative efforts, and these embodiments are also protected by the present invention.

[0031] Example 1: State Monitoring and Prediction of the Servo Cable of the Precision Engraving Robot In this embodiment, the servo cable of a high-precision five-axis linkage CNC engraving robot is subjected to state monitoring and prediction. The servo cables of such robots are subjected to complex bending with high frequency and small curvature radius during operation, and may be eroded by harsh environmental factors such as fine metal dust and cutting fluid vapor generated during the processing, which poses extremely high requirements for the reliability and life prediction of the cables. A method for adaptively predicting the bending stress of the servo cable based on multi-sensor fusion of the present invention can significantly improve the predictive maintenance level of such high-end equipment and effectively reduce the expensive downtime losses and the decline of processing accuracy caused by accidental cable failures.

[0032] Reference Figure 1 The specific implementation steps of the method proposed by the present invention include: obtaining the basic information of the servo cable, the operating parameters of the robot, and collecting the initial historical data of the multi-modal sensor; performing preprocessing operations of online cleaning and spatio-temporal calibration on the above historical data; constructing a deep neural network fusion layer of the entropy weight method and the attention mechanism to fuse the preprocessed data to obtain comprehensive time-frequency domain fusion data; constructing a deep neural network hybrid architecture based on the deep temporal convolutional neural network and the bidirectional long short-term memory neural network to extract features from the fusion data to generate comprehensive time-frequency domain fusion features; performing correlation analysis on the time-frequency domain fusion features through the deep neural network hybrid architecture to obtain a time-domain feature correlation matrix and a frequency-domain feature correlation matrix, and then generating a comprehensive correlation matrix; and obtaining stress prediction data according to the historically acquired bending strain data, temperature gradient data, vibration data, and load current data in real time and the comprehensive correlation matrix.

[0033] First, deploy a multi-modal sensor array at the most severely bent section (about 300 mm in length) inside the X-axis drag chain of the robot. This array includes: 3 distributed fiber Bragg grating sensors FBGs pasted along the axial direction of the cable, with a spacing of 100 mm and a measurement range of ±2500 microstrain , the demodulator outputs the average strain value, and the reference sampling frequency is set to 20 Hz for real-time collection of the bending strain value; 2 micro-infrared array sensors, respectively aimed at the inlet and outlet surfaces of the cable bending section, calculate the average temperature difference between the two points as the gradient value, and the reference sampling frequency is set to 1 Hz; 1 high-sensitivity triaxial MEMS acceleration sensor, model ADXL355, with a range of ±8 g and a bandwidth of 1 kHz, closely attached to the outer sheath of the cable to collect the X, Y, Z-axis vibration acceleration signals as the spectrum value, and the reference sampling frequency is set to 1 kHz; and 1 Hall effect current sensor LEM LTS 6-NP, with a range of ±20 A, connected in series in one phase of the power line of the servo motor to collect the load current value, and the reference sampling frequency is set to 10 kHz. A true stress sensor, used as the label in the offline training stage, is usually a high-precision stress sensor or reference stress data obtained through standard stress-strain conversion.

[0034] All sensor data is synchronously collected through an embedded data acquisition card to form the first historical data set. In a typical training cycle, data with a total duration of 500 hours can be collected. If the average sampling frequency is 100 Hz, the single-modal data volume can reach points.

[0035] Subsequently, online cleaning and spatio-temporal calibration processing are performed on the first historical dataset to generate the second historical dataset. In the online cleaning stage, for the high-frequency noise and transient interference that may be contained in the three-axis acceleration signal sampled at 1 kHz and the load current signal sampled at 10 kHz, five-layer wavelet decomposition is performed using the db4 wavelet basis. For the high-frequency detail coefficients obtained from each layer of decomposition, the noise threshold T is dynamically calculated by the following formula: ; where T is the noise threshold, is the noise standard deviation estimated by dividing the median absolute deviation (MAD) of the high-frequency detail coefficients of the first layer of wavelet decomposition by 0.6745. For a sub-band of a certain acceleration sensor, if its MAD is 0.008 mV, then . N is the number of sampling points in each 1-second data window. For the 1-kHz acceleration signal, N = 1000. At this time, . Then, the soft threshold function is applied to numerically process the wavelet coefficient w, that is, when the wavelet coefficient value is less than the threshold T, it is set to zero, otherwise it is shrunk to effectively suppress noise interference and retain the transient digital characteristics of the strain signal. For the 20-Hz bending strain and 1-Hz temperature gradient signals, the criterion is adopted, that is, if a data point deviates from its local mean by more than three times the local standard deviation, it is regarded as an anomaly and replaced with the mean of adjacent points to eliminate obvious outliers.

[0036] In the spatio-temporal calibration stage, the data records from sensors with different sampling rates are aligned. Taking the time stamp of the 10-kHz current sensor as the benchmark, for the data of the remaining low-sampling-rate sensors, such as 1-Hz temperature gradient, 20-Hz strain, and 1-kHz vibration, cubic spline interpolation algorithm is used for upsampling or alignment to ensure that the time synchronization error is less than or equal to 0.05 milliseconds. For the 1-Hz temperature data points, 10,000 points are generated by cubic spline interpolation between two time stamps to align with the 10-kHz current data. Then, the dimensional difference between different data is eliminated by performing sliding window normalization operation on the digital data. A sliding window with a window length of 10 seconds is used, and the normalization range is set from 0 to 1, and a small protection constant is added to avoid division by zero. The normalization calculation process is as follows: for each data point in the window, its normalized value is equal to subtracting the window minimum value , and then dividing by the window maximum value subtracting the window minimum value plus the protection constant : ; Among them, is each data point within the window, is the minimum value within the window, is the maximum value within the window, is a protection constant to avoid division by zero. After this step, the strain data and temperature data are converted into dimensionless values after normalization and are comparable during feature fusion; for the high-frequency characteristics of the acceleration sensor, an additional 0.1-second short-time window analysis is performed to extract the local statistical features of transient shock events, that is, the peak value, root mean square (RMS), and kurtosis of the acceleration signal are calculated within this 0.1-second window.

[0037] The preprocessing process proposed by the present invention effectively removes noise interference through online cleaning while retaining the transient digital features of key data; the finally saved model through spatio-temporal calibration has the best prediction performance and good generalization ability, enhancing the reliability and accuracy of the prediction method in practical applications.

[0038] Next, based on the second historical dataset, fusion is performed through the constructed deep neural network fusion layer of the entropy weight method and the attention mechanism to obtain comprehensive time-frequency domain fusion data.

[0039] First, calculate the information entropy weight through the entropy weight method. For the bending strain, temperature gradient, vibration, and load current data obtained by each multi-modal sensor in the second historical dataset, divide the value range of the normalized data sequence into 20 equal-width intervals. Calculate the frequency of data points appearing in each interval i (i is the interval, j is the sensor modality), and then calculate the information entropy using the information entropy formula : ; Among them, , m = 20 is the number of bins.

[0040] The information entropy weight is calculated through the following formula: ; Among them, is the number of sensors, that is, four sensors for bending strain, temperature gradient, vibration, and load current.

[0041] If the calculated information entropies of strain, temperature, vibration, and current are 0.88, 0.75, 0.91, and 0.82 respectively, then their corresponding information entropy values are 0.12, 0.25, 0.09, and 0.18 respectively. The sum of these values is 0.12 + 0.25 + 0.09 + 0.18 = 0.64. Therefore, the information entropy weights are: strain temperature , vibration , current .

[0042] Taking the time-frequency domain features of each multi-modal sensor in the second historical dataset as input, that is, the time-domain features of the mean, variance, peak-to-peak value of each multi-modal sensor modality and the frequency-domain features of the energy ratio of each sub-band of the wavelet packet decomposition of the vibration signal, based on the attention mechanism, including a single-layer fully connected neural network layer with 16 hidden units, using the Softmax activation function, output the attention weights representing the feature importance , if after network calculation, the output attention weights are: strain , temperature , vibration , current .

[0043] The final sensor fusion weights are the information entropy weights and the attention weights of the weighted combination, using linear weighted average, and the weight coefficient is 0.5: ; Therefore, the final fusion weight of strain .

[0044] The present invention combines the entropy weight method and the attention mechanism, enables the machine to automatically judge the importance of different sensor data and perform intelligent fusion, avoids the deviation that may be brought by artificially setting fixed weights, thereby obtaining more comprehensive and reliable fusion data, and improves the accuracy of subsequent stress prediction.

[0045] By constructing a trainable deep neural network hybrid architecture based on a deep temporal convolutional neural network and a bidirectional long short-term memory neural network, the comprehensive time-frequency domain fusion data is used for feature extraction to generate comprehensive time-frequency domain fusion features.

[0046] The deep temporal convolutional neural network module consists of three dilated convolutional digital computing units. The size of the convolutional kernel in each layer is fixed at 3×1. Through the dilatation factors that multiply layer by layer, with d = 1 in the first layer, d = 2 in the second layer, and d = 4 in the third layer, convolutional operations are performed on the input time series data with a length of 32 time steps, corresponding to 3.2 seconds of data at a sampling rate of 10 Hz. The bidirectional long short-term memory neural network module is connected after the output of the deep temporal convolutional neural network module. The output sequence of the deep temporal convolutional neural network module serves as the input of the bidirectional long short-term memory neural network module, which contains two independent LSTM layers, one forward and one backward, with 64 hidden layer nodes and a time step of 10. Through gating mechanisms such as input gates, forget gates, output gates, and vector arithmetic operations, the influence of historical and future data on the current predicted value is learned from the forward and backward directions respectively. The output layer of the model is connected after the final output of the bidirectional long short-term memory neural network module, and through a fully connected digital computing structure and the ReLU function, a non-linear activation function, to ensure that the output is non-negative to generate the predicted stress-related data.

[0047] The hybrid architecture of the temporal convolutional network and bidirectional long short-term memory network adopted in the present invention can deeply mine the hidden time laws and long-range dependence relationships in the data, overcome the deficiencies of traditional single models in processing complex time series data, and thus capture and predict the dynamic stress changes of servo cables more accurately.

[0048] Then, the initial learning rate value of the above deep neural network hybrid architecture is set to 0.005. In order to achieve rapid convergence in the initial stage of training and fine-tuning in the later stage, a decaying learning rate strategy is adopted, that is, the learning rate value is adjusted by 5% every 10 training epochs for parameter optimization. The calculation formula is: ; where, is the new learning rate, which is the learning rate value that the model will use in the next training epoch after one decay adjustment. is the current learning rate, which is the learning rate value that the model is currently using before this decay adjustment, that is, the set initial learning rate of 0.005. If the training is carried out for a total of 100 epochs, then at the end of the 10th epoch, the learning rate will become 0.005×0.95 = 0.00475. At the end of the 20th epoch, the learning rate will become 0.00475×0.95≈0.0045125.

[0049] The Huber loss function is selected as the loss function, and the delta parameter is set to 1.0. This function performs piecewise arithmetic operations on the prediction error e. When the absolute value of the error is less than the delta parameter When using the mean squared error calculation, which is greater than or equal to the delta parameter When using the linear error calculation, which combines the calculation of the mean squared error term and the linear error term to evaluate the model performance: ; where, is the delta parameter, which is a key hyperparameter of the Huber loss function and defines the threshold point at which the behavior of the loss function switches; e is the prediction error, which is the difference between the model prediction value and the true value: ; where, is the true stress value, which is the actual observed, sensor-collected, or calibrated true bending stress value of the servo cable. In the training dataset, it is provided to the model as a label (Ground Truth); is the model-predicted stress value, which is the value output by the hybrid architecture of the deep neural network based on the deep temporal convolutional neural network and the bidirectional long short-term memory neural network for predicting the bending stress of the servo cable.

[0050] The Adam algorithm is selected as the optimizer. Adam is an optimization algorithm with an adaptive learning rate. It dynamically adjusts the learning rate of each parameter by calculating the first moment estimate (mean) and the second moment estimate (non-central variance) of the gradient. With momentum parameters beta1 = 0.9 and beta2 = 0.999, the Adam algorithm can update the model parameters, such as the convolutional kernel weights of the TCN and the gating unit weights of the BiLSTM, through the gradient descent algorithm based on backpropagation, thereby accelerating the convergence speed and improving the training efficiency.

[0051] During the training process, the dataset is divided into a training set, a validation set, and a test set in a ratio of 7:1:2. After each training epoch, the Huber loss of the model is evaluated on the validation set. The early stopping strategy is used, that is, based on the comparison logic of the validation set error, if the validation set loss does not decrease for 20 consecutive training epochs, the training is stopped. This effectively prevents the model from overfitting on the training set and ensures that the numerical set of the model parameters with the best performance on the validation set during the training process is finally saved to ensure that the model has good generalization ability.

[0052] The model training method of the present invention can train the model faster and better by dynamically adjusting the learning rate, using the robust Huber loss function, the efficient Adam optimizer, and the early stopping strategy. At the same time, it effectively prevents the model from overfitting the training data and affecting its prediction performance on new data, thereby ensuring that the finally obtained prediction model is both accurate and has good practical application ability.

[0053] Then, calculate the correlations between the time-domain features and frequency-domain features of each original data and the predicted stress, and generate the time-domain feature correlation matrix and the frequency-domain feature correlation matrix.

[0054] Obtain the respective time-domain features and frequency-domain features of the bending strain data, temperature gradient data, vibration data, and load current data from the comprehensive time-frequency domain fusion features through the deep neural network hybrid architecture. Calculate the correlations between the time-domain features and frequency-domain features of each original data and the predicted stress, and generate the time-domain feature correlation matrix and the frequency-domain feature correlation matrix through the Pearson correlation coefficient and the attention mechanism weights inside the model.

[0055] Among them, the Pearson correlation coefficient is calculated for each extracted time-domain feature and frequency-domain feature between its sequence and the corresponding true stress value sequence in the training dataset. The calculation formula of the Pearson correlation coefficient is: ; where X is the feature sequence, Y is the true stress sequence, and are their averages, is the number of data points. Organize the correlation coefficients of all features into the time-domain feature correlation matrix and the frequency-domain feature correlation matrix. If the correlation coefficient between the strain mean and the stress is 0.85, the correlation coefficient between the temperature mean and the stress is 0.30, and the correlation coefficient between the vibration and the stress is 0.75.

[0056] Among them, the weight based on the attention mechanism inside the model means that if the inside of the deep neural network hybrid architecture contains an interpretable attention layer, add an attention module before or after the BiLSTM layer, and directly extract the weights assigned to the time-domain features and frequency-domain features by this attention layer when processing them. These weights directly reflect the degree of attention of the model to specific features during prediction and their importance. Then perform arithmetic mean or weighted mean on the elements in the corresponding positions.

[0057] Among them, the time-domain correlation matrix (simplified to a comprehensive time-domain weight for each sensor) ; the frequency-domain correlation matrix (simplified to a comprehensive frequency-domain weight for each sensor) .

[0058] Finally, perform weighted combination on the time-domain feature correlation matrix and the frequency-domain feature correlation matrix, and obtain the comprehensive correlation matrix by performing arithmetic mean or weighted mean on the elements in the corresponding positions. The comprehensive correlation matrix is obtained through the following operation: ; Among them, the comprehensive correlation matrix of the strain data is calculated as: .

[0059] The method for pattern correlation analysis through a deep neural network hybrid architecture can extract the key time-domain and frequency-domain features of each original data type from complex comprehensive time-frequency domain fusion features, and quantify the correlation strength between these features and the predicted stress by calculating the Pearson correlation coefficient, etc., to construct a time-domain and frequency-domain feature correlation matrix, and finally obtain a comprehensive correlation matrix through weighted combination. At the same time, it provides a certain interpretability for model prediction, avoiding the deficiency of the "black box" operation of deep learning models, thus revealing the internal pattern of the influence of different physical quantity features on cable stress, and providing key knowledge priors for accurate and fast real-time prediction based on this matrix in the future.

[0060] When the system performs real-time stress prediction, it will obtain bending strain data, temperature gradient data, vibration data, and load current data in real time. After preprocessing and preliminary time-frequency domain feature extraction, real-time time-frequency domain feature vectors are obtained. The real-time time-frequency domain feature vectors and the comprehensive correlation matrix are calculated. Through a linear mapping layer L, whose weights have been learned and derived from , the real-time features are mapped into stress feature data: ; Here, L is a simple weighted sum, where the weights come from the contribution of each feature learned in to the predicted stress. Suppose is a vector containing all important time-frequency domain features, including the normalized strain mean, temperature mean, vibration RMS, and current mean, while provides the weight vector corresponding to these features. Then, the stress prediction data is obtained through weighted sum: ; Suppose the real-time time-frequency domain feature vector (after normalization) is , then .

[0061] Among them, Bias is a constant offset. After system calibration, Bias is 50 MPa.

[0062] The method for obtaining stress prediction data can quickly and directly map real-time input to stress prediction values through tensor multiplication of the multi-modal sensor data features obtained and processed in real time with the pre-learned comprehensive correlation matrix. At the same time, it effectively utilizes the established feature-stress correlation pattern, avoiding the need to re-perform complex model inference or feature learning processes during each real-time prediction. Thus, it realizes efficient and real-time servo cable bending stress output, provides a quick response for immediate health status monitoring and early warning decision-making, and enhances the practicality and dynamic adaptability of the system.

[0063] The present invention comprehensively collects cable status information by deploying a multi-modal sensor array, and realizes adaptive and high-precision prediction of the bending stress of the servo cable by combining the intelligent fusion of the entropy weight method and the attention mechanism, the powerful feature extraction ability of the deep time series network, and correlation matrix analysis. This method can comprehensively consider various influencing factors, dynamically adjust the prediction model, improve the robustness and accuracy of the prediction, and provide a decision-making basis for the real-time health monitoring and predictive maintenance of the servo cable.

[0064] Embodiment 2: Bending stress prediction of the servo cable of a single-axis engraving robot This embodiment mainly focuses on a single-axis engraving robot mainly responsible for linear motion. The bending mode of its servo cable is relatively fixed, but it still faces high-frequency bending and environmental factor influences.

[0065] Refer to Figure 2 , the specific implementation steps of the method proposed by the present invention include: obtaining the basic information of the servo cable, the operation parameters of the robot, and the initial historical data collection of the multi-modal sensors; performing preprocessing operations of online cleaning and spatio-temporal calibration on the above historical data; constructing a deep neural network fusion layer of the entropy weight method and the attention mechanism to fuse the preprocessed data to obtain comprehensive time-frequency domain fusion data; constructing a trainable deep neural network hybrid architecture based on a deep time series convolutional neural network and a bidirectional long short-term memory neural network to extract features from the fusion data to generate comprehensive time-frequency domain fusion features; performing correlation analysis on the time-frequency domain fusion features through the deep neural network hybrid architecture to obtain a time-domain feature correlation matrix and a frequency-domain feature correlation matrix, and then generating a comprehensive correlation matrix; obtaining stress prediction data according to the real-time obtained historical bending strain data, temperature gradient data, vibration data, and load current data and the comprehensive correlation matrix; triggering different levels of early warning responses by constructing a health assessment and multi-level early warning module to calculate the cumulative damage of the predicted stress data and according to the degree of damage.

[0066] First, deploy a multi-modal sensor array in the bending section of the servo cable inside the X-axis drag chain of the robot. This array includes: Bending Strain Sensor: Two resistive strain gauges are pasted along the center of the cable outer sheath, with a measurement range of ±1500 microstrain , and the strain value is output through a strain gauge. The reference sampling frequency is set to 15 Hz.

[0067] Temperature Gradient Sensor: One platinum resistance temperature sensor with an accuracy of ±0.5 °C and a response time of 0.5 s, which measures the surface temperature of the cable. The reference sampling frequency is set to 5 Hz.

[0068] Vibration Sensor: One single-axis MEMS acceleration sensor with a range of ±16 g and a bandwidth of 500 Hz, which is closely attached to the cable outer sheath to collect vibration acceleration signals perpendicular to the cable movement direction. The reference sampling frequency is set to 500 Hz.

[0069] Load Current Sensor: One clamp-on Hall effect current sensor with a range of ±400 A and an accuracy of ±1.5%, which is clamped on the main power line of the servo motor to collect the RMS value of the AC load current. The reference sampling frequency is set to 500 Hz.

[0070] True Stress Sensor: Calibrated through a high-precision force sensor on the test bench.

[0071] All sensor data is synchronously collected through an embedded data acquisition card to form the first historical data set. In a typical training cycle, data with a total duration of 200 hours can be collected. If the average sampling frequency is 50 Hz, the single-modal data volume can reach points.

[0072] Subsequently, online cleaning and spatio-temporal calibration processing are performed on the first historical data set to generate the second historical data set. In the online cleaning stage, for the high-frequency noise and transient interference that may be contained in the single-axis acceleration signal sampled at 500 Hz and the load current signal sampled at 500 Hz, four-layer wavelet decomposition is performed using the db4 wavelet basis. For the high-frequency detail coefficients obtained from each layer of decomposition, the noise threshold T is dynamically calculated through the following formula: ; where T is the noise threshold, is the noise standard deviation estimated by dividing the median absolute deviation (MAD) of the high-frequency detail coefficients of the first layer of wavelet decomposition by 0.6745. For a sub-band of an acceleration sensor, for the acceleration signal, N = 500, and after MAD estimation , at this time, . Then, the soft threshold function is applied to numerically process the wavelet coefficients w. For the bending strain at 15 Hz and the temperature data at 5 Hz, median filtering (window size 5 points) is used to smooth the data and remove isolated outliers, rather than strict The criteria are adapted to its lower sampling frequency and relatively stable signal characteristics.

[0073] In the spatio-temporal calibration stage, the data records from sensors with different sampling rates are aligned. Based on the time stamps of the current sensors at 500 Hz, for the data of the remaining low-sampling-rate sensors, such as the 5 Hz temperature gradient and 15 Hz strain, linear interpolation algorithms are used for upsampling or alignment to ensure that the time synchronization error is less than or equal to 0.1 millisecond. For the 5 Hz temperature data, 99 points are generated by linear interpolation between two time stamps and aligned with the 500 Hz current data. Then, the dimensional differences between different data are eliminated by performing a sliding window normalization operation on the digital data. A sliding window with a window length of 5 seconds is used, and the normalization range is set from 0 to 1, with a small protection constant added to avoid division by zero. The normalization calculation process is as follows: for each data point within the window, its normalization is equal to subtracting the minimum value of the window and then dividing by the maximum value of the window subtracting the minimum value of the window plus the protection constant : ; where is each data point within the window, is the minimum value within the window, is the maximum value within the window, is the protection constant to avoid division by zero. After this step, the strain data and temperature data are converted into dimensionless values after normalization and are comparable during feature fusion; for the high-frequency characteristics of the acceleration sensors, an additional 0.1-second short-time window analysis is performed to extract the local statistical features of transient shock events, including calculating the peak value, root mean square, and kurtosis of the acceleration signal within this 0.1-second window.

[0074] Next, based on the second historical dataset, fusion is performed through the constructed deep neural network fusion layer of the entropy weight method and the attention mechanism to obtain the integrated time-frequency domain fusion data.

[0075] First, the information entropy weights are calculated by the entropy weight method. For the data obtained by each multi-modal sensor in the second historical dataset (bending strain, temperature gradient, vibration, load current), the value range of the normalized data sequence is divided into 15 equal-width intervals. Calculate the frequency of data points appearing in each interval i (i is the interval, j is the sensor modality), and then apply the information entropy formula to calculate the information entropy : Among them, , is the number of bins.

[0076] Information entropy weight is calculated by the following formula: ; Among them, is the number of sensors, namely four sensors of bending strain, temperature gradient, vibration, and load current.

[0077] If the calculated strain, temperature, vibration, and current information entropies are 0.80, 0.70, 0.90, and 0.78 respectively, then their corresponding information entropy values are 0.20, 0.30, 0.10, and 0.22 respectively. The sum of these values is 0.20 + 0.30 + 0.10 + 0.22 = 0.82. Therefore, the information entropy weights are: strain , temperature , vibration , current .

[0078] Taking the time-frequency domain features of each multi-modal sensor in the second historical dataset as input, including the time-domain features of the mean, variance, and peak-to-peak value of each multi-modal sensor mode and the frequency-domain features of the energy ratio of each sub-band of the wavelet packet decomposition of the vibration signal, based on the attention mechanism, including a single-layer fully connected neural network layer with 16 hidden units, using the Softmax activation function, output the attention weights representing the feature importance , if after network calculation, the output attention weights are: strain , temperature , vibration , current .

[0079] The final sensor fusion weight is the weighted combination of the information entropy weight and the attention weight , using linear weighted average, and the weight coefficient is 0.5: ; Therefore, the final fusion weight of strain .

[0080] By constructing a trainable deep neural network hybrid architecture based on a deep temporal convolutional neural network and a bidirectional long short-term memory neural network, the comprehensive time-frequency domain fusion data is used for feature extraction to generate comprehensive time-frequency domain fusion features.

[0081] The deep temporal convolutional neural network module consists of 2 dilated convolutional digital computing units. The convolutional kernel size of each layer is fixed at 3×1. Through the dilatation factors that multiply layer by layer, with d = 1 for the first layer and d = 2 for the second layer, convolution operations are performed on the input time series data, which has a length of 20 time steps and corresponds to 1.33 seconds of data at a sampling rate of 15 Hz. The bidirectional long short-term memory neural network module is connected after the output of the deep temporal convolutional neural network module. The output sequence of the deep temporal convolutional neural network module serves as the input to the bidirectional long short-term memory neural network module. It contains two independent LSTM layers, one forward and one backward, with the hidden layer nodes configured to be 32 and the time step length being 8. Through gating mechanisms such as input gates, forget gates, output gates, and vector arithmetic operations, it learns the influence of historical and future data on the current predicted value from both the forward and backward directions respectively. The output layer of the model is connected after the final output of the bidirectional long short-term memory neural network module. Through a fully connected digital computing structure and the non-linear activation function ReLU function, it ensures that the output is non-negative to generate the predicted stress-related data.

[0082] Then, the initial learning rate value of the above deep neural network hybrid architecture is set to 0.002. In order to achieve fast convergence in the initial stage of training and perform fine optimization in the later stage, a decaying learning rate strategy is adopted, that is, the learning rate value is adjusted by 5% every 15 training epochs for parameter optimization. The calculation formula is: ; where, is the new learning rate, which is the learning rate value that the model will use in the next training epoch after one decay adjustment. is the current learning rate, which is the learning rate value that the model is currently using before this decay adjustment, that is, the set initial learning rate of 0.002. If the training is carried out for a total of 100 epochs, then at the end of the 15th epoch, the learning rate will become .

[0083] The Huber loss function is selected as the loss function, with the delta parameter set to 0.8. The Adam algorithm is selected as the optimizer. Adam is an optimization algorithm with an adaptive learning rate. It dynamically adjusts the learning rate of each parameter by calculating the first moment estimate (mean) and second moment estimate (non-centered variance) of the gradient. The momentum parameters are beta1 = 0.9 and beta2 = 0.999. The Adam algorithm can update the model parameters, such as the convolutional kernel weights of TCN and the gating unit weights of BiLSTM, through the gradient descent algorithm based on backpropagation, thereby accelerating the convergence speed and improving the training efficiency.

[0084] During the training process, the dataset is divided into a training set, a validation set, and a test set in a ratio of 7:1:2. After each training epoch, the Huber loss of the model is evaluated on the validation set. An early stopping strategy is used, that is, based on the comparison logic of the validation set error, if the validation set loss does not decrease for 15 consecutive training epochs, the training is stopped. This effectively prevents the model from overfitting on the training set and ensures that the model parameter value set with the best performance on the validation set during the training process is finally saved to ensure that the model has good generalization ability.

[0085] Then, calculate the correlation between the time-domain features and frequency-domain features of each original data and the predicted stress to generate the time-domain feature correlation matrix and the frequency-domain feature correlation matrix.

[0086] Through the deep neural network hybrid architecture, the respective time-domain features and frequency-domain features of the bending strain data, temperature gradient data, vibration data, and load current data are obtained from the comprehensive time-frequency domain fusion features. Calculate the correlation between the time-domain features and frequency-domain features of each original data and the predicted stress, and generate the time-domain feature correlation matrix and the frequency-domain feature correlation matrix through the Pearson correlation coefficient and the attention mechanism weights inside the model.

[0087] Among them, the Pearson correlation coefficient is calculated for each extracted time-domain feature and frequency-domain feature to calculate its Pearson correlation coefficient with the corresponding true stress value sequence in the training dataset.

[0088] Among them, the weights based on the attention mechanism inside the model mean that if the inside of the deep neural network hybrid architecture contains an interpretable attention layer, an attention module is added before or after the BiLSTM layer to directly extract the weights assigned to the time-domain features and frequency-domain features by this attention layer when processing them. These weights directly reflect the degree of attention of the model to specific features and their importance during prediction, and then perform arithmetic mean or weighted mean on the elements in the corresponding positions.

[0089] Among them, the time-domain correlation matrix (simplified to a comprehensive time-domain weight for each sensor) ; the frequency-domain correlation matrix (simplified to a comprehensive frequency-domain weight for each sensor) .

[0090] Finally, the time-domain feature correlation matrix and the frequency-domain feature correlation matrix are weighted and combined, and the comprehensive correlation matrix is obtained by performing arithmetic mean or weighted mean on the elements in the corresponding positions. The comprehensive correlation matrix is obtained through the following operations: ; The calculation for the corresponding strain data is: .

[0091] When the system performs real-time stress prediction, it will obtain bending strain data, temperature gradient data, vibration data, and load current data in real time. After preprocessing and preliminary time-frequency domain feature extraction, a real-time time-frequency domain feature vector is obtained. , the real-time time-frequency domain feature vector and the comprehensive correlation matrix are calculated through a linear mapping layer L, whose weights have been learned and derived from during the training phase, mapping the real-time features to stress feature data: ; Here, L is a simple weighted sum, where the weights come from the contribution of each feature learned from to the predicted stress. Suppose is a vector containing all important time-frequency domain features, including the standardized strain mean, temperature mean, vibration RMS, and current mean, while provides the weight vector corresponding to these features, , then the stress prediction data is obtained through weighted summation: ; After system calibration, the Bias is 40 MPa.

[0092] The system constructs a health assessment and multi-level early warning module to calculate the cumulative damage of the predicted stress data and trigger different levels of early warning responses according to the damage degree. The Miner linear cumulative damage criterion is used to calculate the damage degree D: ; Among them, is the actual number of cycles at a specific stress amplitude level, is the fatigue life at this stress amplitude. The predicted stress interval is divided into 10 levels, and the actual number of cycles at each level is counted.

[0093] Finally, the system triggers different levels of early warning responses according to the damage degree: When the cumulative damage exceeds 60% and is less than or equal to 80% of the threshold , a first-level early warning is triggered: the system log is recorded, and a prompt "40% - 20% of the cable life remaining, it is recommended to conduct a preventive inspection" pops up on the CNC system interface.

[0094] When it is more than 80% and less than or equal to 95%, a secondary warning is triggered: audible and visual alarm (buzzer rings, warning light flashes), and the operating speed of the robot is limited to 70% of the rated value (implemented through robot control interface commands).

[0095] When it is more than 95%, a tertiary warning is triggered: forced shutdown, and a detailed fault diagnosis report is generated and uploaded to the cloud for maintenance personnel to analyze.

[0096] Through the carefully designed sensor deployment, efficient data preprocessing, scientific data fusion, accurate model prediction, flexible dynamic adjustment and reliable warning mechanism in the above embodiments, the real-time monitoring and adaptive prediction of cable stress are comprehensively realized. This effectively improves the service life and reliability of the cable, and has significant practical application value and broad promotion prospects.

[0097] Although specific examples of the present invention have been shown, those of ordinary skill in the art should understand that various adjustments, modifications, substitutions and improvements to these examples are possible without departing from the core principles and spirit of the present invention. The protection scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A servo cable bending stress adaptive prediction method based on multi-sensor fusion, characterized in that, Including: Deploy a multi-modal sensor array in the bending area of the servo cable to obtain a first historical data set, including bending strain data, temperature gradient data, vibration data, load current data, and stress data, and preprocess the first historical data set to obtain a second historical data set; Construct a deep neural network fusion layer based on the entropy weight method and the attention mechanism to fuse the second historical data set to obtain comprehensive time-frequency domain fusion data; Construct a trainable deep neural network hybrid architecture based on the temporal convolutional neural network and the bidirectional long short-term memory neural network to perform deep feature extraction on the comprehensive time-frequency domain fusion data to generate comprehensive time-frequency domain fusion features; Perform pattern correlation analysis on the time-frequency domain fusion features through the deep neural network hybrid architecture to obtain a time-domain feature correlation matrix and a frequency-domain feature correlation matrix, and then obtain a comprehensive correlation matrix; Obtain bending strain data, temperature gradient data, vibration data, and load current data in real time, input them into the deep neural network hybrid architecture, and perform time series prediction inference in combination with the comprehensive correlation matrix to obtain stress prediction data.

2. The adaptive prediction method for the bending stress of a servo cable based on multi-sensor fusion according to claim 1, characterized in that, The preprocessing process includes: Online cleaning processing, using a digital signal filtering method of wavelet threshold denoising to eliminate high-frequency vibration interference, performing multi-layer wavelet decomposition operation on the first historical data set to match the transient digital features of the strain signal, and retaining valid data; among them, the noise threshold value is dynamically calculated according to the length of the digital signal and the noise standard deviation, and the soft threshold function is applied for numerical processing to suppress noise interference; spatio-temporal calibration processing, applying a numerical interpolation alignment algorithm to adjust the data timestamps of sensors with different sampling rates, and eliminating the dimensional difference between different data through sliding window normalization operation.

3. A servo cable bending stress adaptive prediction method based on multi-sensor fusion according to claim 1, characterized in that, The construction of the deep neural network fusion layer based on the entropy weight method and the attention mechanism to fuse the second historical data set includes: Calculate the information entropy weight of the data information entropy obtained by each multi-modal sensor in the second historical data set through the entropy weight method, and the information entropy reflects the information density of the data; use the time-frequency domain features of each sensor in the second historical data set as the input, and output the attention weight matrix representing the feature importance based on the attention mechanism; the final sensor fusion weight is the weighted combination of the information entropy weight and the attention weight to obtain a dynamic fusion weight vector for the adaptive fusion of multi-modal data.

4. A servo cable bending stress adaptive prediction method based on multi-sensor fusion according to claim 1, characterized in that The described trainable deep neural network hybrid architecture is: The temporal convolutional neural network module consists of multiple layers of learnable dilated convolutional kernel parameters, and performs convolutional operations on the input time series data through the dilated factor that doubles layer by layer to achieve coverage of a large receptive field of the time series and capture digital dependencies across different time scales; the bidirectional long short-term memory neural network module is connected after the output of the temporal convolutional neural network module, and its hidden layer nodes are configured with a predetermined number, and learn the influence of historical and future data on the current prediction through the bidirectional temporal backpropagation algorithm; the output layer of the deep neural network hybrid architecture generates the predicted stress-related digital data through a fully connected digital computing structure and a non-linear activation function.

5. A servo cable bending stress adaptive prediction method based on multi-sensor fusion according to claim 4, characterized in that The training process of the deep neural network hybrid architecture includes performing the following adjustments to the model parameters: Adopt a preset initial learning rate value and perform parameter optimization by decaying and adjusting this learning rate value at a predetermined period; select the Huber loss function as the loss function to perform piecewise arithmetic operations on the prediction error, and evaluate the performance of the deep neural network hybrid architecture by combining the calculation of the squared error term and the linear error term; select the Adam algorithm as the optimizer, update the model parameters through the gradient descent algorithm based on backpropagation, and cooperate with the early stopping strategy, that is, the comparison logic based on the validation set error to prevent the deep neural network hybrid architecture from overfitting, and finally save the set of the best model parameter values during the training process.

6. A servo cable bending stress adaptive prediction method based on multi-sensor fusion according to claim 1, characterized in that, The deep neural network hybrid architecture performs pattern correlation analysis on the time-frequency domain fusion features to obtain a time-domain feature correlation matrix and a frequency-domain feature correlation matrix, and then obtains a comprehensive correlation matrix including: The deep neural network hybrid architecture obtains the time-domain features and frequency-domain features of bending strain data, temperature gradient data, vibration data, and load current data from the comprehensive time-frequency domain fusion features; calculates the Pearson correlation coefficients between the time-domain features and frequency-domain features of each original data and the predicted stress, and generates the time-domain feature correlation matrix and the frequency-domain feature correlation matrix; finally, performs weighted combination on the time-domain feature correlation matrix and the frequency-domain feature correlation matrix to obtain the comprehensive correlation matrix.

7. A servo cable bending stress adaptive prediction method based on multi-sensor fusion according to claim 1, characterized in that The method for obtaining stress prediction data is: The bending strain data, temperature gradient data, vibration data, and load current data obtained in real time, after preprocessing and preliminary deep learning feature extraction, are used as the input of the comprehensive correlation matrix, and perform tensor multiplication operation with the comprehensive correlation matrix to generate the final stress prediction data.

Images