An adaptive prediction method for servo cable bending stress based on multi-sensor fusion

Abstract

The present invention relates to the technical field of industrial equipment condition monitoring, and proposes a servo cable bending stress prediction method based on multi-sensor fusion. In view of the problems in the prior art of single data dimension, poor generalization ability of machine learning models and insufficient real-time performance, first, by deploying multi-modal sensors, multi-source historical data sets are acquired and preprocessed. Then, a neural network fusion layer is constructed to fuse the preprocessed data sets to generate comprehensive time-frequency domain fusion data. Next, a trainable deep neural network hybrid architecture is constructed to perform deep feature extraction on the fused data and perform pattern association analysis to obtain a comprehensive association matrix. Finally, the new data acquired in real time is directly input into the prediction neural network, and time series prediction reasoning is performed in combination with the comprehensive association matrix to obtain the final stress prediction data. This method significantly improves the accuracy of servo cable bending stress prediction and enhances the adaptability and robustness of the system.

Description

Technical Field

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

[0002] As a key component of industrial automation equipment, servo cables are susceptible to fatigue, aging, and even breakage under complex mechanical stresses such as frequent bending and torsion, leading to production interruptions, economic losses, and safety hazards. Traditional passive maintenance or periodic replacement strategies are costly and ineffective in preventing sudden failures, making them difficult to meet the high reliability and predictive maintenance requirements of modern industry.

[0003] Existing methods use a single or limited number of sensors combined with shallow neural networks for stress prediction. While these methods are simple to deploy, they lack accuracy under complex operating conditions because they fail to fully account for the combined impact of multi-physics coupling effects on cable stress, such as temperature, load current, and vibration. Furthermore, shallow neural network architectures struggle to fully capture these complex nonlinear relationships and dynamic changes, limiting their effectiveness under variable operating conditions.

[0004] To improve performance, some research has turned to more complex traditional machine learning algorithms, combined with optimization techniques such as gradient descent. While these methods can theoretically optimize model parameters, they typically rely on time-consuming offline training, making them difficult to adapt to the demanding requirements of industrial scenarios for real-time monitoring, millisecond-level response, and dynamic model updates. Their predictive capabilities are particularly limited when faced with sudden changes in operating conditions.

[0005] In recent years, deep learning neural network technology, especially deep time-series convolutional neural networks and bidirectional long short-term memory neural networks, has shown great potential in processing time-series data and extracting complex features. It has significantly improved the accuracy of cable bending stress prediction and can better capture nonlinear dynamics and long-range dependencies. Despite this, there is still room for improvement in the efficiency of feature fusion in current applications. Traditional fusion methods fail to fully explore the deep spatiotemporal 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 opaque and it is difficult to explain the specific impact of each input factor on the prediction results. This, to a certain extent, limits their widespread 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 technology.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A servo cable bending stress adaptive prediction method based on multi-sensor fusion includes:

[0009] A multimodal sensor array is deployed 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 a preprocessing operation of online cleaning and spatiotemporal calibration is performed to obtain a second historical data set;

[0010] Construct a deep neural network fusion layer with entropy weight method and attention mechanism to fuse the second historical data set to obtain comprehensive time-frequency domain fusion data;

[0011] A trainable deep neural network hybrid architecture based on a deep temporal convolutional neural network and a bidirectional long short-term memory neural network is constructed to extract features from the integrated time-frequency domain fusion data and generate integrated time-frequency domain fusion features; correlation analysis is performed on the time-frequency domain fusion features using the neural network architecture to obtain a time domain feature correlation matrix and a frequency domain feature correlation matrix, thereby obtaining a comprehensive correlation matrix;

[0012] Bending strain data, temperature gradient data, vibration data and load current data are acquired in real time, and stress prediction data is obtained based on these four data and the comprehensive correlation matrix.

[0013] Preferably, the pretreatment process includes:

[0014] The first historical data is cleaned online, and a digital filtering method with wavelet threshold denoising is used to eliminate high-frequency vibration interference. A five-layer decomposition is performed using the DB4 wavelet basis. The DB4 wavelet basis is a commonly used compactly supported orthogonal wavelet basis with excellent time-frequency localization properties, making it suitable for multi-scale signal analysis and able to effectively match the transient characteristics of strain signals, thereby preserving the signal's effective information. This method uses a five-layer wavelet decomposition, with the number of decomposition layers selected based on the length and characteristics of the signal. This five-layer decomposition effectively removes high-frequency noise while preserving the signal's main features. However, the specific number of decomposition layers needs to be adjusted according to the specific signal conditions to achieve the best denoising effect.

[0015] This method performs multi-layer wavelet decomposition on the first historical data set to match the transient digital characteristics of the strain signal and retain valid data. The noise threshold is dynamically calculated through parameter optimization based on the length of the digital signal and the noise standard deviation. Wavelet decomposition of the digital signal is performed using a soft threshold function to obtain wavelet coefficients at each scale to suppress noise interference. The soft threshold is calculated based on a universal threshold rule combined with the estimated noise standard deviation of each layer. The specific steps are as follows:

[0016] By performing j-layer wavelet decomposition on the input signal and using the db4 wavelet basis to perform 5-layer decomposition, the detail coefficients of each decomposition layer j (j=1, ..., J) are obtained. . For each decomposition layer j, the detail coefficient , the median absolute deviation (MAD) method is used to estimate the noise standard deviation , noise standard deviation The calculation formula is as follows:

[0017] ;

[0018] in, .

[0019] Combined with an estimate of its noise standard deviation and the length of the detail coefficient of this layer Calculate. The threshold calculation formula for each decomposition layer is:

[0020] ;

[0021] in, is the threshold calculated at the jth decomposition layer, is the estimated value of the noise standard deviation of the detail coefficient of the j-th decomposition layer, is the signal length of the detail coefficient of the j-th decomposition layer.

[0022] The data is then time-space aligned. Numerical interpolation algorithms are used to align digital timestamps of data records from sensors with different sampling rates, such as the temperature gradient sensor with a sampling rate of 1 Hz and the current sensor with a sampling rate of 50 kHz. Cubic spline interpolation is used to ensure that the timestamp alignment error is less than or equal to 0.05 milliseconds. A sliding window normalization operation is performed on the digital data, with a window length of 10 seconds, a normalization range of 0 to 1, and a small protection constant added. To avoid zero division, the dimension difference between different data is eliminated. The normalization calculation process is: for each data point in the window , its normalized value equal Subtract the window minimum Then divide by the maximum value of the window Subtract the window minimum Protection constant sum:

[0023] ;

[0024] in, .

[0025] The strain and temperature data are normalized to the range of [0, 1] and then converted into dimensionless values, making them comparable during feature fusion. In view of the high-frequency characteristics of the accelerometer, an additional 0.1-second short-term window analysis is added to extract the local statistical features of transient impact events.

[0026] Preferably, the step of constructing a deep neural network fusion layer of an entropy weight method and an attention mechanism to fuse the second historical data set includes:

[0027] The entropy weight method is used to calculate the information entropy of the data obtained by each multimodal sensor in the second historical data set to obtain the information entropy weight. The data is divided into 20 intervals and the probability distribution is calculated. (i is the interval, j is the sensor mode), and the information entropy formula is applied to calculate the information entropy of the j-th sensor mode ;

[0028] ;

[0029] in, , m is the number of bins. This k value is used to convert the information entropy Normalized to the interval [0, 1].

[0030] The information entropy weight of the j-th sensor modality It is calculated by the following formula:

[0031] ;

[0032] in, is the total number of sensors, is the information entropy of the k-th sensor modality, and when summing, k traverses all sensors.

[0033] The time-frequency domain features of each multimodal sensor in the second historical dataset are used as input. Based on the attention mechanism, a single-layer fully connected neural network with 16 hidden units is included. The Softmax activation function is used to take the time-domain features of each multimodal sensor, such as mean, variance, and peak-to-peak value, as input and output the attention weight representing the importance of the feature. , which represents the attention weight of the j-th sensor modality, all Form a vector The final sensor fusion weight of the j-th sensor modality is is the information entropy weight of the j-th sensor modality The attention weight of the j-th sensor modality The weighted combination of , using linear weighted average, with a weight coefficient of 0.5:

[0034] ;

[0035] Preferably, the trainable deep neural network hybrid architecture is:

[0036] The deep temporal convolutional neural network module consists of three layers of dilated convolutional neural network units, with the convolution kernel size of each layer fixed at 3×1. The dilation factor is multiplied layer by layer, with the number of layers d=1, 2, and 4. The input time series data covers 32 time steps, corresponding to 3.2 seconds of data at a 10Hz sampling rate. Convolution operations are performed to cover a wide range of receptive fields 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, with a hidden layer node configuration of 64 and a time step of 10. Through gating mechanisms (such as input gates, forget gates, and output gates) and parameter matrix operations, the influence of historical and future data on the current predicted values ​​is learned from forward and reverse directions respectively; the output layer of the model uses a fully connected neural network layer and a nonlinear activation function ReLU function to ensure that the output is a non-negative value to generate the predicted stress-related digital data.

[0037] Preferably, the training process of the model includes:

[0038] The initial learning rate value is set to 0.005, and the parameter optimization is performed by adjusting the learning rate value by 5% every 10 training cycles, that is:

[0039] ;

[0040] in, Is the new learning rate, which is the learning rate value that the model will use in the next training cycle after a decay adjustment. is the current learning rate, which is the learning rate value currently used by the model before this decay adjustment, that is, the initial learning rate set to 0.005. 5% is the learning rate decay rate.

[0041] The loss function uses the Huber loss function, and the delta parameter is set to 1.0. The function performs a segmented arithmetic operation on the prediction error e. When the absolute value of the error Less than the delta parameter When the square error is used, it is greater than or equal to the delta parameter When , linear error calculation is used, which combines the calculation of square error term and linear error term to evaluate model performance:

[0042] ;

[0043] in, is the delta parameter of the Huber loss function, which is a key hyperparameter of the Huber loss function and defines the threshold point at which the loss function behavior switches; e is the prediction error, which is the difference between the model prediction value and the true value:

[0044] ;

[0045] in, is the true stress value, which is the actual bending stress value of the servo cable that is actually observed, collected by the sensor, or calibrated. In the training dataset, it is provided to the model as a label; The stress value predicted by the model is the output value of the servo cable bending stress predicted by the deep neural network hybrid architecture based on deep temporal convolutional neural network and bidirectional long short-term memory neural network.

[0046] The optimizer uses the Adam algorithm with momentum parameters beta1=0.9 and beta2=0.999. Model parameters, such as the convolution kernel weights of the TCN and the gate unit weights of the BiLSTM, are updated using a backpropagation-based gradient descent algorithm. An early stopping strategy is used in conjunction with a comparison logic based on the validation set error to prevent overfitting of the model. Ultimately, the model parameters with the best performance during training are retained.

[0047] Preferably, the correlation analysis of the time-frequency domain fusion features is performed through a deep neural network hybrid architecture to obtain a time-domain feature correlation matrix and a frequency-domain feature correlation matrix, and then the comprehensive correlation matrix is ​​obtained, including:

[0048] The neural network hybrid architecture obtains the time domain features of bending strain data, temperature gradient data, vibration data and load current data, such as mean, variance, peak, kurtosis, and frequency domain features, such as energy spectrum density, main frequency, bandwidth, and frequency domain center of gravity, from the comprehensive time and frequency domain fusion features. The correlation between the time domain features and frequency domain features of each original data and the predicted stress is calculated, and the time domain feature correlation matrix and frequency domain feature correlation matrix are generated through the Pearson correlation coefficient and the neural network self-attention weight. Among them, the Pearson correlation coefficient The calculation formula is:

[0049] ;

[0050] in, and Represents the i-th observation value of the two variables, and They represent the means of the two variables, and N represents the number of observations.

[0051] Neural network self-attention weight: used to capture the complex relationship between features, the calculation formula is:

[0052] ;

[0053] Among them, Q, K and V represent query vector, key vector and value vector respectively. represents the dimension of the key vector, and the softmax function is used to convert the correlation into a probability distribution.

[0054] Finally, the time domain feature correlation matrix and frequency domain feature correlation matrix Perform weighted combination and obtain the comprehensive correlation matrix by taking arithmetic average or weighted average of the elements in the corresponding positions , the calculation formula is:

[0055] ;

[0056] in, is the weighting coefficient and can be adjusted according to actual needs. The value range of is [0,1]. When , the comprehensive correlation matrix is ​​completely determined by the frequency domain feature correlation matrix; when When , the comprehensive correlation matrix is ​​completely determined by the time domain feature correlation matrix; when When , the comprehensive correlation matrix is ​​determined by the time domain feature correlation matrix and the frequency domain feature correlation matrix.

[0057] Preferably, the method for obtaining stress prediction data based on the four data and the comprehensive correlation matrix is:

[0058] The real-time acquired bending strain data, temperature gradient data, vibration data and load current data are preprocessed and preliminarily extracted from the time-frequency domain features to obtain the real-time time-frequency domain feature vector. , the real-time time-frequency domain feature vector With the comprehensive correlation matrix , calculated through the neural network weight mapping layer, through a linear mapping layer L, whose weights have been trained from Learning derivation maps real-time features into stress feature data:

[0059] ;

[0060] The final stress prediction data is generated after calculation.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] 1. Through the deployment of multimodal sensors and data fusion, the present invention can obtain more comprehensive servo cable status information. Combined with online cleaning and spatiotemporal calibration processing based on a neural network preprocessing layer, the quality and consistency of input data are significantly improved, laying a solid foundation for subsequent accurate prediction.

[0063] 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 weight according to the information volume and importance of each sensor data, thereby enhancing the effectiveness of deep neural network feature extraction and its adaptability to changes in working conditions; 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.

[0064] 3. Through correlation analysis of deep neural networks, the present invention can quantify and generate feature correlation matrices and comprehensive correlation matrices in the time domain and frequency domain, clearly revealing the specific contributions of different sensor data dimensions to stress prediction, enhancing the interpretability of the prediction model, and providing direct weight guidance for real-time stress prediction.

[0065] 4. The present invention uses the comprehensive correlation matrix learned by the neural network to directly guide the stress prediction of real-time multimodal data, achieving real-time and accurate prediction, effectively supporting the predictive maintenance of servo cables, and reducing failure risks and operation and maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0067] Figure 1 A flow chart showing the state monitoring and prediction of the servo cable of a precision engraving robot proposed in an embodiment of the present invention;

[0068] Figure 2 This is a flowchart of the state monitoring and prediction of the servo cable of a single-axis engraving robot proposed in an embodiment of the present invention. DETAILED DESCRIPTION

[0069] Next, the technical solution of the present invention will be explained in detail and clearly in conjunction with specific examples of the present invention and related drawings. It should be understood that the examples described below represent only a part of the present invention, not all of it. Based on these examples, a person of ordinary skill in the art can deduce other possible implementations of the present invention without making any creative efforts, and these methods are also protected by the present invention.

[0070] Example 1: Condition Monitoring and Prediction of Servo Cables for Precision Engraving Robots

[0071] This embodiment monitors and predicts the condition of the servo cable of a high-precision five-axis CNC engraving robot. The servo cable of such a robot is subjected to complex bending with high frequency and small curvature radius during operation. At the same time, it may be corroded by harsh environmental factors such as fine metal dust and cutting fluid vapor generated during the processing, which places extremely high demands on the reliability and life prediction of the cable. The present invention's adaptive prediction method for servo cable bending stress based on multi-sensor fusion can significantly improve the predictive maintenance level of such high-end equipment, effectively reducing expensive downtime losses and reduced processing accuracy caused by unexpected cable failures.

[0072] Reference Figure 1 The specific implementation steps of the method proposed in the present invention include: obtaining basic information of the servo cable, robot operating parameters and initial historical data collection of multimodal sensors; performing preprocessing operations of online cleaning and spatiotemporal calibration on the above historical data; constructing a deep neural network fusion layer with entropy weight method and attention mechanism to fuse the preprocessed data to obtain comprehensive time-frequency domain fusion data; constructing a deep neural network hybrid architecture based on deep temporal convolutional neural network and bidirectional long short-term memory neural network to extract features from the fused data and 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 generate a comprehensive correlation matrix; obtaining stress prediction data based on the real-time acquired historical bending strain data, temperature gradient data, vibration data and load current data and the comprehensive correlation matrix.

[0073] First, a multimodal sensor array was deployed at the section of the robot's X-axis drag chain where the servo cable is most severely bent (approximately 300 mm in length). The array consisted of three distributed fiber grating (FBG) sensors attached along the cable axis, with a spacing of 100 mm and a measurement range of ±2500 microstrain. The demodulator outputs average strain values, with a base sampling frequency of 20 Hz, for real-time acquisition of bending strain values. Two miniature infrared array sensors, aimed at the entrance and exit surfaces of the cable bend, calculate the average temperature difference between the two points as the gradient value, with a base sampling frequency of 1 Hz. A high-sensitivity triaxial MEMS accelerometer (ADXL355) with a range of ±8 g and a bandwidth of 1 kHz is placed in close contact with the cable outer sheath, collecting X, Y, and Z triaxial vibration acceleration signals as spectral values, with a base sampling frequency of 1 kHz. A Hall-effect current sensor (LEMLTS 6-NP) with a range of ±20 A is connected in series to one phase of the servo motor's power line to collect load current values, with a base sampling frequency of 10 kHz. A real stress sensor is used for labeling during the offline training phase. It is typically a high-precision stress sensor or reference stress data obtained through standard stress-strain conversion.

[0074] All sensor data are collected synchronously through the embedded data acquisition card to form the first historical data set. In a typical training cycle, a total of 500 hours of data can be collected. If the average sampling frequency is 100Hz, the amount of single-modal data can reach point.

[0075] Subsequently, the first historical data set was cleaned and time-space calibrated online to generate a second historical data set. During the online cleaning phase, a five-layer wavelet decomposition was performed using the db4 wavelet basis to remove high-frequency noise and transient interference that may be contained in the 1kHz sampled triaxial acceleration signal and the 10kHz sampled load current signal. For each layer of high-frequency detail coefficients obtained from the decomposition, the noise threshold T was dynamically calculated using the following formula:

[0076] ;

[0077] Where T is the noise threshold, The noise standard deviation is estimated by dividing the median absolute deviation (MAD) of the high-frequency detail coefficients of the first wavelet decomposition by 0.6745. For a sub-band of an accelerometer, its MAD is 0.008mV, then N is the number of sampling points per 1 second data window. For a 1kHz acceleration signal, N=1000. Then, apply the soft threshold function The wavelet coefficient w is numerically processed, that is, when the wavelet coefficient value When the signal is less than the threshold value 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 20Hz bending strain and 1Hz temperature gradient signal, the The criterion is that if a data point deviates from its local mean by more than three times the local standard deviation, it is considered abnormal and replaced by the mean of the adjacent points, and obvious outliers are eliminated.

[0078] During the spatiotemporal calibration phase, data records from sensors with different sampling rates are aligned. Taking the 10kHz current sensor timestamp as the benchmark, the remaining low-sampling-rate sensor data, such as 1Hz temperature gradient, 20Hz strain, and 1kHz vibration, are upsampled or aligned using a cubic spline interpolation algorithm to ensure that the time synchronization error is less than or equal to 0.05 milliseconds. For 1Hz temperature data points, 10,000 points are generated between the two timestamps using cubic spline interpolation to align with the 10kHz current data. Then, a sliding window normalization operation is performed on the digital data to eliminate the dimensional differences between different data. A sliding window with a window length of 10 seconds is used, the normalization range is set to 0 to 1, and a small protection constant is added. To avoid division by zero. The normalization calculation process is: for each data point in the window , its normalized value equal Subtract the window minimum Then divide by the maximum value of the window Subtract the window minimum Protection constant sum:

[0079] ;

[0080] in, is each data point within the window, is the minimum value in the window, is the maximum value within the window, This is a protection constant to avoid division by zero. After this step, the strain and temperature data are normalized and converted to dimensionless values, making them comparable during feature fusion. To address the high-frequency characteristics of the accelerometer, an additional 0.1-second short-term window analysis is added to extract the local statistical features of the transient impact event. Specifically, the peak value, root mean square (RMS), and kurtosis of the acceleration signal are calculated within this 0.1-second window.

[0081] The preprocessing process proposed in the present invention effectively removes noise interference through online cleaning while retaining the transient digital characteristics of key data; the model finally saved through spatiotemporal calibration has the best prediction performance and good generalization ability, enhancing the reliability and accuracy of the prediction method in practical applications.

[0082] Next, based on the second historical data set, the deep neural network fusion layer constructed with the entropy weight method and attention mechanism is fused to obtain comprehensive time-frequency domain fusion data.

[0083] First, the information entropy weight is calculated by the entropy weight method. The range of the normalized data sequence of the bending strain, temperature gradient, vibration, and load current data obtained by each multimodal sensor in the second historical data set is divided into 20 intervals of equal width. The frequency of occurrence of the data point in each interval i is calculated. (i is the interval, j is the sensor mode), and then the information entropy formula is applied to calculate the information entropy :

[0084] ;

[0085] in, , m=20 is the number of bins.

[0086] Information entropy weight It is calculated by the following formula:

[0087] ;

[0088] in, is the number of sensors, namely, bending strain, temperature gradient, vibration, and load current sensors.

[0089] If the calculated information entropy of strain, temperature, vibration, and current are 0.88, 0.75, 0.91, and 0.82 respectively, then the corresponding information entropy is The 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: ,temperature ,vibration , current .

[0090] The time-frequency domain features of each multimodal sensor in the second historical data set are used as input, namely the time-domain features of the mean, variance, and peak-to-peak value of each multimodal sensor mode and the frequency-domain features of the energy proportion of each sub-band of the wavelet packet decomposition of the vibration signal. Based on the attention mechanism, a single-layer fully connected neural network layer with 16 hidden units is included, and the Softmax activation function is used to output the attention weight representing the importance of the feature. , if calculated by the network, the output attention weight is: strain ,temperature ,vibration , current .

[0091] Final sensor fusion weights is the information entropy weight With the attention weight The weighted combination of , using linear weighted average, with a weight coefficient of 0.5:

[0092] ;

[0093] Therefore, the final fusion weight of the strain .

[0094] By combining the entropy weight method and the attention mechanism, the present invention enables the machine to automatically determine the importance of different sensor data and perform intelligent fusion, avoiding the deviation that may be caused by manually setting fixed weights, thereby obtaining more comprehensive and reliable fusion data and improving the accuracy of subsequent stress prediction.

[0095] 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 subjected to feature extraction to generate comprehensive time-frequency domain fusion features.

[0096] The deep temporal convolutional neural network module consists of three layers of dilated convolution digital computing units, with the convolution kernel size of each layer fixed at 3×1. Through the dilation factor doubling layer by layer, the first layer d=1, the second layer d=2, and the third layer d=4, the input time series data with a length of 32 time steps, corresponding to 3.2 seconds of data at a 10Hz sampling rate, is subjected to convolution operations; the bidirectional long short-term memory neural network module is connected after the output of the deep temporal convolutional neural network module, and 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. It contains two independent LSTM layers, one forward and one reverse, with a hidden layer node configuration of 64 and a time step of 10. Through gating mechanisms such as input gate, forget gate, output gate and vector arithmetic operations, it learns the influence of historical and future data on the current predicted value from the forward and reverse 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 a nonlinear activation function ReLU function, ensures that the output is non-negative to generate the predicted stress-related data.

[0097] The hybrid architecture of the time series convolutional network and the bidirectional long short-term memory network adopted in the present invention can deeply mine the hidden temporal patterns and long-range dependencies in the data, overcoming the shortcomings of the traditional single model in processing complex time series data, thereby more accurately capturing and predicting the dynamic stress changes of the servo cable.

[0098] 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 early 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 decayed by 5% every 10 training cycles. The calculation formula is:

[0099] ;

[0100] in, is the new learning rate, which is the learning rate value that the model will use in the next training cycle after a decay adjustment. is the current learning rate. This is the initial learning rate of 0.005, which was used by the model before this decay adjustment. Training is performed for 100 epochs. At the end of the 10th epoch, the learning rate will be 0.005 × 0.95 = 0.00475. At the end of the 20th epoch, the learning rate will be 0.00475 × 0.95 ≈ 0.0045125.

[0101] The loss function uses the Huber loss function, and the delta parameter is set to 1.0. The function performs a segmented arithmetic operation on the prediction error e. When the absolute value of the error Less than the delta parameter When the square error is used, it is greater than or equal to the delta parameter When , linear error calculation is used, which combines the calculation of square error term and linear error term to evaluate model performance:

[0102] ;

[0103] in, is the delta parameter, which is a key hyperparameter of the Huber loss function and defines the threshold point at which the loss function behavior switches; e is the prediction error, which is the difference between the model prediction value and the true value:

[0104] ;

[0105] in, is the true stress value, which is the actual bending stress value of the servo cable that is actually observed, collected by the sensor, or calibrated. In the training dataset, it is provided to the model as a label (Ground Truth); The stress value predicted by the model is the output value of the servo cable bending stress predicted by the deep neural network hybrid architecture based on deep temporal convolutional neural network and bidirectional long short-term memory neural network.

[0106] The Adam algorithm is used as the optimizer. It is an adaptive learning rate optimization algorithm that dynamically adjusts the learning rate of each parameter by calculating the first-order moment estimate (mean) and second-order moment estimate (uncentered variance) of the gradient. Momentum parameters beta1=0.9 and beta2=0.999. The Adam algorithm updates model parameters, such as the convolution kernel weights of the TCN and the gate unit weights of the BiLSTM, through a backpropagation-based gradient descent algorithm, thereby accelerating convergence and improving training efficiency.

[0107] During training, the dataset was split into training, validation, and test sets in a ratio of 7:1:2. After each training epoch, the model's Huber loss was evaluated on the validation set. An early stopping strategy, based on a comparison of the validation set's error, was used. Training was stopped if the validation set loss did not decrease within 20 consecutive training epochs. This effectively prevented overfitting of the model on the training set and ensured that the model parameter values ​​that achieved the best performance on the validation set were ultimately retained, ensuring good generalization of the model.

[0108] The model training method of the present invention can make the model training faster and better by dynamically adjusting the learning rate, using the robust Huber loss function, the efficient Adam optimizer and the early stopping strategy, while effectively preventing the model from overfitting the training data and affecting its prediction performance on new data, thereby ensuring that the final prediction model is both accurate and has good practical application capabilities.

[0109] Then, the correlation between the time domain features and frequency domain features of each original data and the predicted stress is calculated to generate the time domain feature correlation matrix and the frequency domain feature correlation matrix.

[0110] The deep neural network hybrid architecture is used to extract the time and frequency domain features of bending strain data, temperature gradient data, vibration data, and load current data from the integrated time and frequency domain fusion features. The correlation between the time and frequency domain features of each raw data and the predicted stress is calculated. The time and frequency domain feature correlation matrices are generated using the Pearson correlation coefficient and the weights of the attention mechanism within the model.

[0111] Among them, the Pearson correlation coefficient is calculated for each extracted time domain feature and frequency domain feature between its corresponding true stress value sequence in the training data set. The calculation formula of the Pearson correlation coefficient is:

[0112] ;

[0113] Where X is the characteristic sequence, Y is the true stress sequence, and is its average value, is the number of data points. The correlation coefficients of all features are organized into a time domain feature correlation matrix and a frequency domain feature correlation matrix. For example, if the correlation coefficient between mean strain and stress is 0.85, the correlation coefficient between mean temperature and stress is 0.30, and the correlation coefficient between vibration and stress is 0.75.

[0114] The model-based attention mechanism weights represent the weights of a deep neural network hybrid architecture that includes an interpretable attention layer. Adding an attention module before or after the BiLSTM layer directly extracts the weights assigned by the attention layer when processing time and frequency domain features. These weights directly reflect the model's focus and importance on specific features during prediction. The elements at corresponding positions are then arithmetic averaged or weighted averaged.

[0115] Among them, the time domain correlation matrix (simplified to a comprehensive time domain weight for each sensor) ; Frequency domain correlation matrix (simplified to a comprehensive frequency domain weight for each sensor) .

[0116] Finally, the time domain feature correlation matrix and the frequency domain feature correlation matrix are weightedly combined, and the comprehensive correlation matrix is ​​obtained by performing arithmetic averaging or weighted averaging on the elements at corresponding positions. Obtained through the following operations:

[0117] ;

[0118] The comprehensive correlation matrix of strain data is calculated as: .

[0119] The method of pattern association analysis using a deep neural network hybrid architecture can extract 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 predicted stress by calculating the Pearson correlation coefficient, etc., to achieve the construction of 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 degree of interpretability for model prediction, avoids the shortcomings of the "black box" operation of the deep learning model, and thus reveals the inherent pattern of the influence of different physical quantity characteristics on cable stress, providing key knowledge priors for subsequent accurate and rapid real-time prediction based on this matrix.

[0120] 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, the real-time time-frequency domain feature vector is obtained. , the real-time time-frequency domain feature vector With the comprehensive correlation matrix , is calculated by passing through a linear mapping layer L, whose weights have been obtained from the training phase Learning derivation maps real-time features into stress feature data:

[0121] ;

[0122] Here L is a simple weighted sum, where the weights come from The contribution of each feature learned in to the predicted stress. Assume is a vector that contains all important time-frequency domain features, including the normalized strain mean, temperature mean, vibration RMS, and current mean, while Provides the weight vectors corresponding to these features , , then the stress prediction data By weighted summation we get:

[0123] ;

[0124] Assume that the real-time time-frequency domain feature vector (normalized) is ,but .

[0125] Among them, Bias is a constant offset. After system calibration, Bias is 50MPa.

[0126] The method for obtaining stress prediction data can quickly and directly map real-time input to stress prediction values ​​by performing tensor multiplication operations on multimodal sensor data features acquired and processed in real time and a pre-learned comprehensive correlation matrix. This method effectively utilizes the established feature-stress correlation pattern, avoiding the need to re-perform complex model reasoning or feature learning processes for each real-time prediction, thereby achieving efficient and real-time servo cable bending stress output, providing a rapid response for immediate health status monitoring and early warning decision-making, and enhancing the practicality and dynamic adaptability of the system.

[0127] This method deploys a multimodal sensor array to comprehensively collect cable status information. Combining the intelligent fusion of entropy weighting and attention mechanisms, the powerful feature extraction capabilities of deep temporal networks, and correlation matrix analysis, it achieves adaptive and highly accurate prediction of servo cable bending stress. This method comprehensively considers multiple influencing factors, dynamically adjusts the prediction model, and improves the robustness and accuracy of predictions, providing a decision-making basis for real-time health monitoring and predictive maintenance of servo cables.

[0128] Example 2: Bending stress prediction of servo cable of a single-axis engraving robot

[0129] This embodiment is mainly aimed at a single-axis engraving robot that is mainly responsible for linear motion. The bending mode of its servo cable is relatively fixed, but it is still affected by high-frequency bending and environmental factors.

[0130] Reference Figure 2 The specific implementation steps of the method proposed in the present invention include: obtaining basic information of the servo cable, robot operating parameters and initial historical data collection of multimodal sensors; performing preprocessing operations of online cleaning and spatiotemporal calibration on the above historical data; constructing a deep neural network fusion layer with entropy weight method and 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 deep temporal convolutional neural network and bidirectional long short-term memory neural network to extract features from the fused data and 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 generate a comprehensive correlation matrix; obtaining stress prediction data based on the real-time historical bending strain data, temperature gradient data, vibration data and load current data and the comprehensive correlation matrix; by constructing a health assessment and multi-level warning module, calculating the cumulative damage of the predicted stress data and triggering different levels of warning responses according to the degree of damage.

[0131] First, a multimodal sensor array is deployed on the curved section of the servo cable inside the robot's X-axis drag chain. The array includes:

[0132] Bending strain sensor: 2 resistance strain gauges attached along the center of the cable outer sheath, with a measurement range of ±1500 microstrain , the strain value is output through the strain gauge, and the reference sampling frequency is set to 15Hz.

[0133] Temperature gradient sensor: 1 platinum resistance temperature sensor, accuracy ±0.5℃, response time 0.5 seconds, measures the cable surface temperature, and the base sampling frequency is set to 5Hz.

[0134] Vibration sensor: 1 single-axis MEMS accelerometer with a range of ±16g and a bandwidth of 500Hz. It is placed close to the outer sheath of the cable and collects vibration acceleration signals perpendicular to the direction of cable movement. The base sampling frequency is set to 500Hz.

[0135] Load current sensor: 1 clamp-type Hall effect current sensor with a range of ±400A and an accuracy of ±1.5%. It is clamped on the main power line of the servo motor and collects the RMS value of the AC load current. The base sampling frequency is set to 500Hz.

[0136] Real stress sensor: Calibrated by high-precision load cell on the test bench.

[0137] All sensor data are collected synchronously through the embedded data acquisition card to form the first historical data set. In a typical training cycle, a total of 200 hours of data can be collected. If the average sampling frequency is 50Hz, the amount of single-modal data can reach point.

[0138] Subsequently, the first historical data set was cleaned and time-space calibrated online to generate a second historical data set. During the online cleaning phase, a four-layer wavelet decomposition was performed using the db4 wavelet basis to remove high-frequency noise and transient interference that may be contained in the 500Hz sampled uniaxial acceleration signal and the 500Hz sampled load current signal. For each layer of high-frequency detail coefficients obtained from the decomposition, the noise threshold T was dynamically calculated using the following formula:

[0139] ;

[0140] Where T is the noise threshold, The noise standard deviation is estimated by dividing the median absolute deviation (MAD) of the high-frequency detail coefficients of the first wavelet decomposition by 0.6745. For a sub-band of an acceleration sensor, for acceleration signals, N=500, and MAD estimation ,at this time, Then, a soft threshold function is applied to numerically process the wavelet coefficients w. For the 15 Hz bending strain and 5 Hz temperature data, a median filter (window size 5 points) is used to smooth the data and remove isolated outliers, rather than a strict criterion to adapt to its lower sampling frequency and relatively stable signal characteristics.

[0141] During the spatiotemporal calibration phase, data records from sensors with different sampling rates are aligned. Taking the 500Hz current sensor timestamp as the benchmark, the remaining low-sampling-rate sensor data, such as 5Hz temperature gradient and 15Hz strain, are upsampled or aligned using a linear interpolation algorithm to ensure that the time synchronization error is less than or equal to 0.1 milliseconds. For 5Hz temperature data, 99 points are generated between the two timestamps by linear interpolation and aligned with the 500Hz current data. Then, a sliding window normalization operation is performed on the digital data to eliminate the dimensional differences between different data. A sliding window with a window length of 5 seconds is used, the normalization range is set to 0 to 1, and a small protection constant is added. To avoid division by zero. The normalization calculation process is: for each data point in the window , its normalized equal Subtract the window minimum Then divide by the maximum value of the window Subtract the window minimum Protection constant sum:

[0142] ;

[0143] in, is each data point within the window, is the minimum value in the window, is the maximum value within the window, This is a protection constant to avoid division by zero. After this step, the strain and temperature data are normalized and converted to dimensionless values, making them comparable during feature fusion. To address the high-frequency characteristics of the accelerometer, an additional 0.1-second short-term window analysis is performed to extract the local statistical features of the transient impact event. This includes calculating the peak value, root mean square (RMS), and kurtosis of the acceleration signal within this 0.1-second window.

[0144] Next, based on the second historical data set, the deep neural network fusion layer constructed with the entropy weight method and attention mechanism is fused to obtain comprehensive time-frequency domain fusion data.

[0145] First, the information entropy weight is calculated using the entropy weight method. The range of the normalized data sequence of the data (bending strain, temperature gradient, vibration, load current) obtained by each multimodal sensor in the second historical data set is divided into 15 intervals of equal width. The frequency of occurrence of the data point in each interval i is calculated. (i is the interval, j is the sensor mode), and then the information entropy formula is applied to calculate the information entropy :

[0146] ;

[0147] in, , is the number of bins.

[0148] Information entropy weight It is calculated by the following formula:

[0149] ;

[0150] in, is the number of sensors, namely, bending strain, temperature gradient, vibration, and load current sensors.

[0151] If the calculated information entropy of strain, temperature, vibration, and current are 0.80, 0.70, 0.90, and 0.78 respectively, then the corresponding information entropy is The 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: ,temperature ,vibration , current .

[0152] The time-frequency domain features of each multimodal sensor in the second historical data set are used as input, including the time-domain features of the mean, variance, and peak-to-peak value of each multimodal sensor mode and the frequency-domain features of the energy proportion of each sub-band of the wavelet packet decomposition of the vibration signal. Based on the attention mechanism, a single-layer fully connected neural network layer with 16 hidden units is included, and the Softmax activation function is used to output the attention weight representing the importance of the feature. , if calculated by the network, the output attention weight is: strain ,temperature ,vibration , current .

[0153] Final sensor fusion weights is the information entropy weight With the attention weight The weighted combination of , using linear weighted average, with a weight coefficient of 0.5:

[0154] ;

[0155] Therefore, the final fusion weight of the strain .

[0156] 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 subjected to feature extraction to generate comprehensive time-frequency domain fusion features.

[0157] The deep temporal convolutional neural network module consists of two layers of dilated convolution digital computing units, with the convolution kernel size of each layer fixed at 3×1. The first layer d=1, the second layer d=2, and the input time series data with a length of 20 time steps, corresponding to 1.33 seconds of data at a 15Hz sampling rate, are subjected to convolution operations; the bidirectional long short-term memory neural network module is connected after the output of the deep temporal convolutional neural network module, and 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. It contains two independent LSTM layers, one forward and one reverse, with a hidden layer node configuration of 32 and a time step of 8. Through gating mechanisms such as input gate, forget gate, output gate and vector arithmetic 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 is connected after the final output of the bidirectional long short-term memory neural network module, and a fully connected digital computing structure and a nonlinear activation function ReLU function are used to ensure that the output is non-negative to generate the predicted stress-related data.

[0158] Then, the initial learning rate of the above-mentioned deep neural network hybrid architecture is set to 0.002. In order to achieve rapid convergence in the early 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 decayed by 5% every 15 training cycles. The calculation formula is:

[0159] ;

[0160] in, is the new learning rate, which is the learning rate value that the model will use in the next training cycle after a decay adjustment. Is the current learning rate, which is the learning rate value currently used by the model before this decay adjustment, that is, the initial learning rate set to 0.002. If the training is carried out for a total of 100 cycles, then at the end of the 15th cycle, the learning rate will become .

[0161] The Huber loss function was used as the loss function, with the delta parameter set to 0.8. The Adam algorithm was used as the optimizer. Adam is an adaptive learning rate optimization algorithm that dynamically adjusts the learning rate of each parameter by calculating the first-order moment estimate (mean) and second-order moment estimate (non-centered variance) of the gradient. Momentum parameters beta1=0.9 and beta2=0.999 were used. The Adam algorithm updates model parameters, such as the convolution kernel weights of the TCN and the gate unit weights of the BiLSTM, through a backpropagation-based gradient descent algorithm, thereby accelerating convergence and improving training efficiency.

[0162] During training, the dataset was split into training, validation, and test sets in a ratio of 7:1:2. After each training epoch, the model's Huber loss was evaluated on the validation set. An early stopping strategy, based on a comparison of the validation set's error, was used. Training was stopped if the validation set loss did not decrease within 15 consecutive training epochs. This effectively prevented overfitting of the model on the training set and ensured that the model parameter values ​​that achieved the best performance on the validation set were ultimately retained, ensuring good generalization of the model.

[0163] Then, the correlation between the time domain features and frequency domain features of each original data and the predicted stress is calculated to generate the time domain feature correlation matrix and the frequency domain feature correlation matrix.

[0164] The deep neural network hybrid architecture is used to extract the time and frequency domain features of bending strain data, temperature gradient data, vibration data, and load current data from the integrated time and frequency domain fusion features. The correlation between the time and frequency domain features of each raw data and the predicted stress is calculated. The time and frequency domain feature correlation matrices are generated using the Pearson correlation coefficient and the weights of the attention mechanism within the model.

[0165] The Pearson correlation coefficient is calculated for each extracted time domain feature and frequency domain feature, and the Pearson correlation coefficient between the feature and the corresponding true stress value sequence in the training data set.

[0166] The model-based attention mechanism weights represent the behavior of a deep neural network hybrid architecture that includes an interpretable attention layer. By adding an attention module before or after the BiLSTM layer, we can directly extract the weights assigned by the attention layer when processing time-domain and frequency-domain features. These weights directly reflect the model's focus and importance on specific features during prediction. The elements at the corresponding positions are then arithmetic averaged or weighted averaged.

[0167] Among them, the time domain correlation matrix (simplified to a comprehensive time domain weight for each sensor) ; Frequency domain correlation matrix (simplified to a comprehensive frequency domain weight for each sensor) .

[0168] Finally, the time domain feature correlation matrix and the frequency domain feature correlation matrix are weightedly combined, and the comprehensive correlation matrix is ​​obtained by performing arithmetic averaging or weighted averaging on the elements at corresponding positions. Obtained through the following operations:

[0169] ;

[0170] The calculation of the corresponding strain data is: .

[0171] 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, the real-time time-frequency domain feature vector is obtained. , the real-time time-frequency domain feature vector With the comprehensive correlation matrix , is calculated by passing through a linear mapping layer L, whose weights have been obtained from the training phase Learning derivation maps real-time features into stress feature data:

[0172] ;

[0173] Here L is a simple weighted sum, where the weights come from The contribution of each feature learned in to the predicted stress. Assume is a vector that contains all important time-frequency domain features, including the normalized strain mean, temperature mean, vibration RMS, and current mean, while Provides the weight vectors corresponding to these features , , then the stress prediction data By weighted summation we get:

[0174] ;

[0175] After system calibration, the Bias is 40MPa.

[0176] The system builds a health assessment and multi-level warning module to calculate the cumulative damage of the predicted stress data and trigger different levels of warning responses based on the degree of damage. The damage degree D is calculated using the Miner linear cumulative damage criterion:

[0177] ;

[0178] in, is the actual number of cycles at a specific stress amplitude level, is the fatigue life under this stress amplitude. The predicted stress range is divided into 10 levels, and the actual number of cycles under each level is counted.

[0179] Finally, the system triggers different levels of early warning responses based on the severity of the damage:

[0180] When the cumulative damage exceeds the threshold When the cable life is 60% and less than or equal to 80%, a first-level warning is triggered: the system log is recorded, and a prompt "The cable life is 40% to 20% remaining, and preventive inspection is recommended" pops up on the CNC system interface.

[0181] When it exceeds 80% and is less than or equal to 95%, a secondary warning is triggered: an audible and visual alarm (buzzer rings, warning light flashes), and the robot's operating speed is limited to 70% of the rated value (achieved through the robot control interface command).

[0182] When it exceeds 95%, a third-level warning is triggered: forced shutdown, and a detailed fault diagnosis report is generated and uploaded to the cloud for maintenance personnel to analyze.

[0183] The above examples, through carefully designed sensor deployment, efficient data preprocessing, scientific data fusion, precise model prediction, flexible dynamic adjustment, and reliable early warning mechanisms, fully implement real-time monitoring and adaptive prediction of cable stress. This effectively improves the service life and reliability of cables, possessing significant practical application value and broad prospects for promotion.

[0184] Although specific examples of the present invention have been described, it should be understood by those skilled in the art that various adjustments, modifications, substitutions, and improvements to these examples are possible without violating the core principles and spirit of the present invention. The scope of protection 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: include: A multimodal sensor array is deployed in a 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 the first historical data set is preprocessed to obtain a second historical data set; Construct a deep neural network fusion layer based on entropy weight method and 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 temporal convolutional neural network and bidirectional long short-term memory neural network to perform deep feature extraction on the integrated time-frequency domain fusion data and generate integrated time-frequency domain fusion features; The time-frequency domain fusion features are subjected to pattern association analysis by the deep neural network hybrid architecture to obtain a time-domain feature association matrix and a frequency-domain feature association matrix, and then a comprehensive association matrix is ​​obtained; the pattern association analysis to obtain a time-domain feature association matrix and a frequency-domain feature association matrix, and then a comprehensive association matrix specifically comprises: the deep neural network hybrid architecture obtains the time-domain features and the frequency-domain features of the bending strain data, the temperature gradient data, the vibration data, and the load current data from the comprehensive time-frequency domain fusion features; the Pearson correlation coefficient between the time-domain features and the frequency-domain features of each original data and the predicted stress is calculated to generate the time-domain feature association matrix and the frequency-domain feature association matrix; and finally, the time-domain feature association matrix and the frequency-domain feature association matrix are weightedly combined to obtain the comprehensive association matrix; Bending strain data, temperature gradient data, vibration data, and load current data are acquired in real time, input into the deep neural network hybrid architecture, and combined with the comprehensive correlation matrix to perform time series prediction reasoning to obtain stress prediction data; the method for obtaining the stress prediction data is as follows: the bending strain data, temperature gradient data, vibration data, and load current data acquired in real time are preprocessed and subjected to preliminary deep learning feature extraction, and then used as input to the comprehensive correlation matrix, and tensor multiplication operation is performed with the comprehensive correlation matrix to generate final stress prediction data.

2. The method for adaptively predicting bending stress of a servo cable based on multi-sensor fusion according to claim 1 is characterized in that: The pretreatment process includes: In the online cleaning process, a digital signal filtering method with wavelet threshold denoising is used to eliminate high-frequency vibration interference. Multi-layer wavelet decomposition operations are performed on the first historical data set to match the transient digital characteristics of the strain signal and retain valid data. Among them, the noise threshold value is dynamically calculated based on the length of the digital signal and the noise standard deviation, and a soft threshold function is applied for numerical processing to suppress noise interference. In the spatiotemporal calibration process, a numerical interpolation alignment algorithm is applied to adjust the data timestamps of sensors with different sampling rates, and a sliding window normalization operation is performed to eliminate the dimensional differences between different data.

3. The method for adaptively predicting bending stress of a servo cable based on multi-sensor fusion according to claim 1, characterized in that: The step of constructing a deep neural network fusion layer based on the entropy weight method and the attention mechanism to fuse the second historical data set includes: The information entropy of the data obtained by each multimodal sensor in the second historical data set is calculated using the entropy weight method to obtain the information entropy weight, which reflects the information density of the data; the time-frequency domain features of each sensor in the second historical data set are used as input, and the attention weight matrix representing the importance of the features is output based on the attention mechanism; the final sensor fusion weight is a weighted combination of the information entropy weight and the attention weight, and a dynamic fusion weight vector is obtained for the adaptive fusion of multimodal data.

4. The method for adaptively predicting bending stress of a servo cable based on multi-sensor fusion according to claim 1, characterized in that: The trainable deep neural network hybrid architecture is: The temporal convolutional neural network module is composed of multiple layers of learnable dilated convolution kernel parameters. It performs convolution operations on the input time series data through layer-by-layer multiplication of the dilation factor to achieve coverage of a wide range of receptive fields of the time series, and is used to 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 to a predetermined number. It learns the impact of historical and future data on current predictions through a bidirectional temporal back propagation 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 nonlinear activation function.

5. The method for adaptively predicting bending stress of a servo cable based on multi-sensor fusion according to claim 4 is characterized in that: The training process of the hybrid deep neural network architecture involves adjusting the model parameters as follows: A preset initial learning rate value is used, and the parameter optimization is performed by attenuating the learning rate value according to a predetermined period; the loss function uses the Huber loss function to perform piecewise arithmetic operations on the prediction error, and the calculation of the square error term and the linear error term is combined to evaluate the performance of the deep neural network hybrid architecture; the optimizer uses the Adam algorithm to update the model parameters through a gradient descent algorithm based on backpropagation, and uses an early stopping strategy, that is, a comparison logic based on the validation set error to prevent the deep neural network hybrid architecture from overfitting, and finally saves the model parameter value set with the best performance during the training process.

Images