Stress sensing intelligent prediction method based on deep learning

Abstract

The invention provides a stress sensing intelligent prediction method based on deep learning, and aims to solve the defects of a traditional method in the aspects of force value prediction precision and time sequence modeling capability. The method comprises the following steps: acquiring force value data of a stress luminescent material in a stretching process and image data of a corresponding luminescent area, and further realizing time alignment of the force value data and the image data through a data preprocessing and time alignment matching method so as to construct standardized time sequence data; then, a convolutional neural network based on a gating circulation unit is constructed, the convolutional neural network is used for extracting image features, the gating circulation unit network is responsible for modeling a time relation in time sequence data, and meanwhile, the capturing capacity for a complex stress change rule is further enhanced in combination with a multi-head attention mechanism; and finally, the dynamic change trend of the force value is predicted, a force value prediction result is output, and high-precision and intelligent technical support is provided for industrial monitoring and structural health diagnosis.

Description

Technical Field

[0001] This invention relates to the fields of deep learning and sensor data processing technology, and in particular to a deep learning-based intelligent prediction method for stress sensing. Background Technology

[0002] Stress sensing prediction is one of the important technologies in industrial monitoring and structural health diagnosis. Its core lies in analyzing the trend of stress change over time and its spatial distribution, providing a scientific basis for equipment safety assessment and structural health diagnosis. However, existing stress prediction methods mainly rely on traditional data modeling and statistical analysis methods, which are difficult to fully capture the spatiotemporal variation of complex stress characteristics, especially in dynamic and nonlinear environments where they exhibit insufficient prediction accuracy.

[0003] With the rapid development of deep learning technology, it has demonstrated great potential in the field of complex data modeling and prediction. For example, convolutional neural networks have significant advantages in extracting spatial features from images, while gated recurrent unit networks perform well in processing the temporal dependencies of time series data. Combining these two networks in stress prediction problems can effectively uncover the feature correlation between force data and image data. However, existing models still face problems such as insufficient representation of multimodal data and difficulty in capturing complex stress patterns in feature fusion and temporal modeling applications.

[0004] Currently, multimodal data modeling methods suffer from insufficient efficiency in feature extraction and fusion when applied to stress prediction tasks. For example, the dynamic changes and correlations of multimodal features such as force data and image data in complex environments are not adequately modeled, resulting in traditional methods being unable to accurately predict stress change trends and distribution patterns, thus limiting their application in industrial monitoring and health diagnosis.

[0005] Based on this, the present invention proposes a deep learning-based intelligent prediction method for stress sensing, which constructs a convolutional neural network based on gated recurrent units to improve multimodal data processing capabilities, accurately capture stress change trends, and provide efficient and intelligent support for industrial monitoring and structural health diagnosis. Summary of the Invention

[0006] This invention addresses the shortcomings of existing stress sensing prediction methods in terms of accuracy, robustness, and temporal modeling capabilities by proposing a deep learning-based intelligent stress sensing prediction method. By constructing a convolutional neural network deep learning model based on gated recurrent units and introducing a multi-head attention mechanism, the method effectively enhances the system's ability to capture complex stress variation characteristics, achieving high-precision prediction of stress trends over time. This provides intelligent technical support for industrial monitoring and structural health diagnosis.

[0007] This invention is achieved through the following technical solution:

[0008] Step 1: Record the changes in force values ​​and image data of the luminescent material during uniaxial tension.

[0009] Step 2: Preprocess the force data and image data. The preprocessed force data and image data are synchronized and paired using a time alignment matching method to construct standardized time series data. Then, the time series data is divided into training set and test set in a 7:3 ratio.

[0010] Step 3: Based on the training set from Step 2, construct a convolutional neural network based on gated recurrent units, establish a standardized mapping relationship between time series data and force value data, and complete the model training process;

[0011] Step 4: Input the time series data of the test set into the trained model to predict the dynamic trend of force value changes and output the predicted force value results.

[0012] In step two, the preprocessing of the force data includes: using a time series filtering method to remove high-frequency noise from the force data, and simultaneously removing outlier data whose errors exceed three times the standard deviation of the mean, to obtain cleaned force data; further, using a maximum-minimum normalization method, the cleaned force data is linearly mapped to the interval between 0 and 1, finally obtaining preprocessed force data of size 100×1, where 100 represents the number of force data sampling points and 1 represents that each sampling point has one value; the preprocessing of the image data includes: using Gaussian filtering to eliminate random noise in the image; and cropping the image to a fixed size of 312×75, and further normalizing the image pixel values, finally obtaining preprocessed image data of size 312×75×3, where 312 represents the image height, 75 represents the width, and 3 represents the RGB color channels.

[0013] The time alignment matching process refers to aligning the time point corresponding to the greatest force value in the preprocessed force value data with the time point corresponding to the image data of the brightest luminous region; for image data at other time points, spline interpolation is used to calculate and generate corresponding force value data, thereby constructing standardized time series data including force value time series data and image time series data.

[0014] In step three, the convolutional neural network architecture based on gated recurrent units includes a convolutional neural network and a gated recurrent unit network. The convolutional neural network is used to extract image features such as shape contour, brightness gradient and color distribution from image time series data. The image features and force value time series data are further concatenated and stitched together using a unified dimension method to obtain one-dimensional time series data, which is then input into the gated recurrent unit network for modeling.

[0015] The convolutional neural network includes two convolutional layers, two pooling layers, an activation function layer, and a flattening layer. Specifically, the first convolutional layer has 16 output channels, the second convolutional layer has 32 output channels, and each convolutional layer is equipped with a 2×2 max pooling layer and a ReLU activation function.

[0016] The gated recurrent unit network includes a forward-gated recurrent unit network module, a backward-gated recurrent unit network module, a bidirectional splicing module, and a multi-head attention mechanism module. The forward-gated recurrent unit network module analyzes the trend of force value changes in one-dimensional time series data from the past to the current time point to obtain time series data one. The backward-gated recurrent unit network module analyzes the trend of force value changes in one-dimensional time series data from the future to the current time point to obtain time series data two. The bidirectional splicing module then concatenates time series data one and time series data two using a unified dimension method to obtain the amplitude information of force value changes over time. The multi-head attention mechanism module compares the amplitude information of force value changes over time, uses the Softmax algorithm to normalize the amplitude information at each time point, calculates the weight of each time point, and thus optimizes the attention allocation of the model.

[0017] Based on existing technology, the present invention has the following beneficial technical effects:

[0018] To address the shortcomings of traditional methods in stress sensing data modeling and time-series prediction accuracy, this invention proposes a deep learning-based intelligent stress sensing prediction method. By constructing a convolutional neural network based on gated recurrent units, it achieves efficient modeling of complex stress variation characteristics. Compared with existing methods, this invention can integrate multimodal features of force data and image data, accurately capture the trend of stress changes over time, and improve the modeling accuracy and prediction reliability of complex mechanical behaviors.

[0019] Furthermore, this invention supports the generation of intuitive visualization results, providing a high-precision and intelligent decision support tool for industrial monitoring and structural health diagnosis. Moreover, the model structure of this invention is simple, has a wide range of applications, and has strong versatility and promotional value. Attached Figure Description

[0020] Figure 1 This is a flowchart of the stress sensing intelligent prediction method of the present invention;

[0021] Figure 2 This is a graph showing the change in the accuracy of the model force value in this invention;

[0022] Figure 3 This is a scatter plot comparing the predicted force value and the actual force value of the present invention;

[0023] Figure 4This is the histogram of the predicted force error distribution of the present invention;

[0024] Figure 5 This is a diagram illustrating the implementation steps of the present invention; Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention.

[0026] like Figure 1 As shown, this invention provides a deep learning-based intelligent prediction method for stress sensing. This method mainly includes the following modules: a data acquisition module, a data preprocessing module, a deep learning model module, and a prediction output module.

[0027] Step 1: In the data acquisition module, the stress sensor records the changes in force values ​​of the stress-luminescent material during uniaxial tension. However, in actual acquisition, the force values ​​may have non-fixed sampling frequency characteristics, which may affect the continuity and temporal modeling of the data. Therefore, the force values ​​will be standardized by preprocessing methods such as time interpolation and normalization to enhance their temporal alignment and consistency, and meet the needs of deep learning modeling.

[0028] Meanwhile, by using a precision imaging device to record the changes in image data of the luminescent area during the stretching process of the stress-luminescent material at a frequency of 1 / 60 second, the image data can reflect the temporal characteristics of the optical properties of the measured area as a function of force, thus supplementing the lack of spatial characteristic information in single force value data.

[0029] Step 2: Preprocess the force data and image data. The preprocessed force data and image data are synchronized and paired using a time alignment matching method to construct standardized time series data. Then, the time series data is divided into training set and test set in a 7:3 ratio.

[0030] The preprocessing of the force data includes: using time series filtering to remove high-frequency noise from the force data, and removing outliers with errors exceeding three times the standard deviation of the mean, to obtain cleaned force data; further, using the maximum-minimum normalization method, the cleaned force data is linearly mapped to the interval between 0 and 1, finally obtaining preprocessed force data of size 100×1, where 100 represents the number of force data sampling points and 1 represents that each sampling point has 1 value;

[0031] Specifically, the max-min normalization method refers to normalizing the force data to make it suitable for the input requirements of the deep learning model, using the following formula: Where x represents the original force value data, x max and x min , respectively, represent the minimum and maximum values ​​of the force data, and x′ represents the normalized force data. Normalization ensures that the force data is within the range of 0 to 1.

[0032] The image data preprocessing includes: using Gaussian filtering to eliminate random noise in the image; cropping the image to a fixed size of 312×75; further normalizing the image pixel values; and finally obtaining preprocessed image data with a size of 312×75×3, where 312 represents the height of the image, 75 represents the width, and 3 represents the RGB color channel.

[0033] To ensure the temporal alignment and consistency of force data and image data, this invention proposes a timeline alignment and matching method. The specific process is as follows: The time point corresponding to the greatest force value in the preprocessed force data is aligned with the time point corresponding to the image data of the brightest luminous region. For image data at other time points, spline interpolation is used to calculate and generate the corresponding force data, thereby constructing standardized image and force time series data. The matching formula is: in In the formula F interp (t) represents the interpolated force value, corresponding to time t, F i t represents the i-th force value in the original force data. i Representation of force value data F i The corresponding time point, t j Representation of force value data F j The corresponding time points, where j≠i,ω i represents the weighting coefficient used for smooth interpolation, and n represents the number of sampling points for the original force value data.

[0034] Through the above preprocessing operations, the force data and image data are aligned in time and space, providing high-quality and standardized data input for subsequent multimodal data fusion and deep learning modeling, effectively reducing the impact of noise and inconsistency of the original data on the model training accuracy.

[0035] Step 3: Based on the training set from Step 2, construct a convolutional neural network based on gated recurrent units, establish a standardized mapping relationship between time series data and force value data, and complete the model training process; for example... Figure 2This is a graph showing the change in the model force accuracy of the present invention; it demonstrates the convergence trend of the model over 50 training cycles. The overall prediction accuracy of the model gradually improves and tends to stabilize, eventually converging to a range of approximately 90%-93%, fully demonstrating the model's ability to predict force values ​​under complex mechanical changes.

[0036] The convolutional neural network architecture based on gated recurrent units includes a convolutional neural network and a gated recurrent unit network. The convolutional neural network is used to extract image features such as shape contour, brightness gradient and color distribution from image time series data. The image features and force value time series data are further concatenated and stitched together using a unified dimension method to obtain one-dimensional time series data, which is then input into the gated recurrent unit network for modeling.

[0037] The convolutional neural network comprises two convolutional layers, two pooling layers, an activation function layer, and a flattening layer. Specifically, the first convolutional layer has 16 output channels and a 3×3 kernel size, while the second convolutional layer has 32 output channels and a 3×3 kernel size. Each convolutional layer is equipped with a 2×2 max pooling layer and a ReLU activation function. The pooling layer reduces the size of the feature map while retaining key features to reduce computational complexity. The activation function introduces non-linearity to enhance the model's expressive power, effectively preventing the vanishing gradient problem and accelerating network training convergence. The expression for the activation function is: In the formula, x represents the input value, specifically the output of the convolutional layer; the further flattening layer flattens the output of the convolutional layer into a one-dimensional vector, which is used as the output of image features.

[0038] The method of concatenating and stitching through a unified dimension refers to the following: the one-dimensional image features extracted by the convolutional neural network correspond one-to-one with the force value time series data at each time step, and the image features and force value data at each time point are directly stitched together to obtain one-dimensional time series data, which not only retains the spatial information in the image features, but also includes the numerical characteristics of the force value changing over time.

[0039] The gated recurrent unit network includes a forward-gated recurrent unit network module, a backward-gated recurrent unit network module, a bidirectional stitching module, and a multi-head attention mechanism module. The forward-gated recurrent unit network module analyzes the trend of force value changes in one-dimensional time series data from the past to the current time point to obtain time series data one. The backward-gated recurrent unit network module analyzes the trend of force value changes in one-dimensional time series data from the future to the current time point to obtain time series data two. The bidirectional stitching module then concatenates time series data one and time series data two using a unified dimension method to obtain the amplitude information of force value changes over time. The multi-head attention mechanism module compares the amplitude information of force value changes over time, uses the Softmax algorithm to normalize the amplitude information of each time point, calculates the weight of each time point, and thus optimizes the attention allocation of the model.

[0040] The Softmax algorithm is used to calculate the weights of each time point in the time series data, thereby optimizing the model's prediction of force changes. Specifically, the Softmax algorithm converts the magnitude of the force change at each time point into a probability weight, representing the degree of influence of that time point on the model's prediction result. The magnitude of the force at each time point is exponentialized and normalized to finally calculate the weight. The formula for the Softmax algorithm is: In the formula a i It represents the magnitude of the force at time point i, and n is the length of the time series; The weighted score obtained by exponentializing the magnitude information of the force value at time point i is used to enhance the weight of time points with larger changes; the denominator part This is used to normalize the weights of all time points so that the sum is 1, ensuring that the model can adjust the influence of time points according to their relative importance.

[0041] Step 4: Input the time series data of the test set into the trained model to predict the dynamic trend of force value changes and output the predicted force value results; the overall steps are as follows. Figure 5 The diagram shows the implementation steps of the present invention; Figure 3 The results show a scatter plot comparison between the model's predicted force values ​​and the actual force values. The scatter plot distribution closely approximates the ideal fitted line (dashed line), reflecting the model's high accuracy in predicting the dynamic changes in force values. Figure 4 The histogram shows the distribution of the predicted force error. As can be seen from the figure, the distribution of the prediction error is mainly concentrated around 0, showing a trend close to a normal distribution. This demonstrates the high accuracy and stability of the model in force prediction, with a small error fluctuation range.

[0042] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the patent and are not intended to limit it. For those skilled in the art, several modifications and improvements can be made without departing from the principle of this patent, and these should also be considered within the scope of protection of this patent.

Claims

1. A deep learning-based intelligent prediction method for stress sensing, characterized in that, The method includes the following steps: Step 1: Record the changes in force values ​​and image data of the luminescent material during uniaxial tension. Step 2: Preprocess the force data and image data. The preprocessed force data and image data are synchronized and paired using a time alignment matching method to construct standardized time series data. Then, the time series data is divided into training set and test set in a 7:3 ratio. Step 3: Based on the training set from Step 2, construct a convolutional neural network based on gated recurrent units, establish a standardized mapping relationship between time series data and force value data, and complete the model training process; Step 4: Input the time series data of the test set into the trained model to predict the dynamic trend of force value changes and output the predicted force value results.

2. The stress sensing intelligent prediction method according to claim 1, characterized in that, In step two, the preprocessing of the force data includes: using a time series filtering method to remove high-frequency noise from the force data, and simultaneously removing outlier data whose errors exceed three times the standard deviation of the mean, to obtain cleaned force data; further, using a maximum-minimum normalization method, the cleaned force data is linearly mapped to the interval between 0 and 1, finally obtaining preprocessed force data of size 100×1, where 100 represents the number of force data sampling points and 1 represents that each sampling point has one value; the preprocessing of the image data includes: using Gaussian filtering to eliminate random noise in the image; and cropping the image to a fixed size of 312×75, and further normalizing the image pixel values, finally obtaining preprocessed image data of size 312×75×3, where 312 represents the image height, 75 represents the width, and 3 represents the RGB color channels.

3. The stress sensing intelligent prediction method according to claim 1, characterized in that, In step two, the time alignment matching process refers to aligning the time point corresponding to the greatest force value in the preprocessed force value data with the time point corresponding to the image data of the brightest luminous area. For image data at other time points, spline interpolation is used to calculate and generate corresponding force data, thereby constructing standardized time series data that includes force time series data and image time series data.

4. The stress sensing intelligent prediction method according to claim 1, characterized in that, In step three, the convolutional neural network architecture based on gated recurrent units includes a convolutional neural network and a gated recurrent unit network. The convolutional neural network is used to extract image features such as shape contour, brightness gradient and color distribution from image time series data. The image features and force value time series data are further concatenated and stitched together using a unified dimension method to obtain one-dimensional time series data, which is then input into the gated recurrent unit network for modeling.

5. The stress sensing intelligent prediction method according to claim 4, characterized in that, The convolutional neural network includes two convolutional layers, two pooling layers, an activation function layer, and a flattening layer. Specifically, the first convolutional layer has 16 output channels, the second convolutional layer has 32 output channels, and each convolutional layer is equipped with a 2×2 max pooling layer and a ReLU activation function.

6. The stress sensing intelligent prediction method according to claim 4, characterized in that, The gated loop unit network includes a forward gated loop unit network module, a backward gated loop unit network module, a bidirectional splicing module, and a multi-head attention mechanism module. The forward gated loop unit network module obtains time series data one by analyzing the trend of force value change in one-dimensional time series data from the past to the current time point. The backward gated loop unit network module obtains time series data two by analyzing the trend of force value change in one-dimensional time series data from the future to the current time point. The bidirectional splicing module then concatenates time series data one and time series data two using a unified dimension method to obtain the amplitude information of force value change over time. The multi-head attention mechanism module compares the magnitude information of force values ​​changing over time, uses the Softmax algorithm to normalize the magnitude information at each time point, calculates the weight at each time point, and thus optimizes the attention allocation of the model.

Images