Method for predicting residual life of terahertz photoconductive transmitting antenna
By using the Internet of Things and deep learning algorithms to extract features and predict time series of sensor data from terahertz photoconductive transmitting antennas, the problem of predicting the life of terahertz photoconductive transmitting antennas was solved, and the reliability and operational stability of the equipment were improved.
Patent Information
- Application Number
- CN202510908496.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-02
- Publication Date
- 2025-09-30
AI Technical Summary
The existing technology lacks an effective method to predict the lifespan of terahertz photoconductive transmitting antennas, which makes them easily damaged in abnormal working environments or hardware anomalies, increasing maintenance costs and reducing equipment reliability.
The Internet of Things technology is used to monitor the terahertz time-domain spectroscopy system in real time. The asymmetric convolutional neural network and Transformer algorithm are used to extract features and predict time series of sensor data to evaluate the remaining service life of the transmitting antenna.
The accurate prediction of the remaining life of the terahertz photoconductive emission antenna is achieved, ensuring the reliable operation of the equipment and avoiding the maintenance costs and equipment failures caused by antenna damage.
Smart Images
Figure CN120724847A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of life prediction, and in particular relates to a method for predicting the remaining life of a terahertz photoconductive transmitting antenna. Background Art
[0002] Existing terahertz time-domain spectroscopy systems consist of a bias source, a laser, an optical delay device, a transmitting antenna, a detection antenna, a signal acquisition system, and a display screen. The laser outputs two femtosecond pulses of the same frequency and phase. One femtosecond pulse (pump light) is output to the transmitting antenna, while the other femtosecond pulse (probe light) is delayed by an optical fiber delay device before reaching the detection antenna. The femtosecond pulses excite photogenerated carriers within the photoconductive material of the transmitting antenna. Under the influence of an external bias source, these carriers rapidly migrate and radiate high-repetition-rate pulsed terahertz waves into space. These pulsed terahertz waves propagate to the detection antenna, where they meet the femtosecond pulse (probe light), generating a current. The signal acquisition system samples this current to determine the terahertz signal intensity, which is then processed and displayed on a display screen. The terahertz photoconductive transmitting antenna, used to generate terahertz waves, is the most critical component in a terahertz time-domain spectroscopy system. Generally, proper operation of a terahertz transmitting antenna requires a specific operating environment and stable hardware parameters. If these system components malfunction or the operating environment experiences unusual changes, they can easily damage the system's transmitting antenna, resulting in a failure to generate terahertz waves, high repair costs, and reduced device reliability. Therefore, in actual use, it is necessary to estimate and analyze the lifespan of photoconductive transmitting antennas. Existing technologies do not offer such solutions. Summary of the Invention
[0003] The present invention aims to design a method for predicting the remaining useful life of a terahertz photoconductive transmitting antenna. This method uses Internet of Things (IoT) technology to monitor the status of various components and the operating environment in a terahertz time-domain spectroscopy system in real time. Sensor data from each component is imported into a backend database. An asymmetric convolutional neural network is used to extract features from this sensor data. A Transformer algorithm is then used to perform time series prediction on the extracted features. The method then promptly provides feedback to the terahertz time-domain spectroscopy system regarding potential anomalies and remaining useful life. This deep learning algorithm is used to assess the remaining useful life (RUL) of the transmitting antenna, ensuring reliable device operation.
[0004] In order to achieve the above objectives, the present invention adopts the following technical solutions.
[0005] A method for predicting the remaining lifetime of a terahertz photoconductive emission antenna includes the following steps.
[0006] (1) Data acquisition: Turn on the terahertz time-domain spectroscopy system and collect sensor data when the transmitting antenna works continuously under different working conditions. The working conditions are divided according to the state of the delay line. The sensor data include the temperature of the semiconductor saturable absorber of the femtosecond laser, the temperature of the femtosecond laser chamber, the repetition rate difference of the femtosecond laser, the output power of the femtosecond laser, the temperature of the femtosecond laser pumps 1-3, the current of the femtosecond laser pumps 1-3, the bias source voltage, the working environment temperature of the terahertz time-domain spectroscopy system, the working environment humidity of the terahertz time-domain spectroscopy system, and the vibration limit state data of the terahertz time-domain spectroscopy system.
[0007] Assume that the timing data of the antenna operation is , the life cycle is , the sensor data dimension is , then the timing data of any transmitting antenna under a certain working condition can be expressed as X.
[0008]
[0009] (2) Data processing: First, the K-means algorithm is used to process the sensor data. Clustering is performed, and then normalization is performed based on min-max. The time sliding window is used to divide the data set into multiple time windows to obtain data set samples under different working conditions. The data set is divided into training set, validation set and test set. Finally, each sample is RUL labeled.
[0010] (3) The training set data is input into the CNN-Transformer model for model training, the model parameters are adjusted through the validation set, and the test set data is input into the trained CNN-Transformer model for result prediction. During the process, the asymmetric convolutional neural network is used to extract features from the sensor data after data processing in step (2), and the Transformer algorithm is used to perform time series prediction on the extracted features.
[0011] Step (3) specifically includes the following steps: (301) the asymmetric convolutional neural network includes the first convolutional layer, the second convolutional layer and the third convolutional layer, the first channel of the first convolutional layer uses a 1×3 convolution kernel to extract the relevant features between different time nodes in the time window, the second channel uses a 3×1 convolution kernel to extract the features between the sensor data in the time window, and the third channel uses a 3×3 convolution kernel to extract the data relationship between different time nodes and adjacent data in the time window, and then uses formula (5) to superimpose the results of the 1×3 convolution kernel, the 3×1 convolution kernel and the 3×3 convolution kernel. In the process, the ReLU self-normalization activation function and the batch normalization method are used.
[0012]
[0013] in, is the feature map after superposition of the first convolutional layer, They correspond to the asymmetric convolution feature maps of channels 1 to 3 respectively, and c is the number of corresponding channels.
[0014] (302) The second convolutional layer processes the output of the first convolutional layer in the same way as in step (301) to produce the output feature map. The third convolutional layer uses the same method as step (301) to process the output of the second convolutional layer and output the feature map , and finally connect the feature maps of the three convolutional layers, as shown in formula (6).
[0015]
[0016] in, is the feature map output after connection.
[0017] (303) Feature map First, it passes through the Flatten layer and is converted into a one-dimensional feature vector, and then it is connected to the Liner layer, and the feature vector is fixed to Dimension, and then position encoding. Position encoding is used to add the location information of the data, and the encoding method uses sine and cosine functions, such as equations (7) and (8).
[0018]
[0019]
[0020] Where, is the time step of the window (i.e., sliding step S), is any vector in the time step, is the dimension of the feature vector.
[0021] (304) Position coding will be added The dimensional feature vector is used as the input feature vector Y and input into the encoder layer. The encoder layer consists of a multi-head attention mechanism, a first layer normalization, a first residual connection, a feedforward neural network, a second layer normalization, a second residual connection and a linear regression layer. The encoder layer learns the knowledge mapping to the predicted value of RUL , calculated as in formula (13).
[0022]
[0023] in, is the input of the linear regression layer, is the weight matrix, is the bias vector.
[0024] Step (2) specifically includes the following steps: (201) Considering the complexity of the sensor data of the terahertz time-domain spectroscopy system under different working conditions, in order to eliminate the influence of the operating conditions on the original data sequence, the sensor data is firstly processed by the K-means algorithm. Clustering is performed and then normalization is performed based on min-max, as shown in formula (2).
[0025]
[0026] in, 1 to A positive integer between 1 to A positive integer between and They are The maximum and minimum values in .
[0027] After normalization, the sensor data X is obtained , for.
[0028]
[0029] (202) Figure 3 As shown, a time sliding window is used to Processing is performed to reflect the attenuation time relationship of the transmitting antenna of different sensing signals in the process of terahertz wave generation. The sliding window length is set to , sliding step length , the total sampling length is , the data Split into A time window.
[0030] (203) A segmented method is used to mark the lifetime of the transmitting antenna samples, as shown in formula (4). Under normal circumstances, The antenna will not experience signal attenuation or the signal has very little attenuation within the time range. After a period of time, the antenna will experience significant signal attenuation. Based on the current experience of using the transmitting antenna, set It is 700h.
[0031] .
[0032] Compared with existing technologies, this invention offers the following advantages: it utilizes an asymmetric convolutional neural network to extract features from sensor data, employs a Transformer algorithm to perform time series prediction on these extracted features, and promptly provides feedback to the terahertz time-domain spectroscopy system regarding potential anomalies and remaining lifespan. This deep learning algorithm enables assessment of the remaining useful life (RUL) of the transmitting antenna, ensuring reliable device operation. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 The present invention is a flowchart of a method for predicting the remaining lifetime of a terahertz photoconductive transmitting antenna.
[0034] Figure 2 This is a schematic diagram of the CNN-Transformer model structure involved in the present invention.
[0035] Figure 3 Schematic diagram of the time sliding window partitioning process.
[0036] Figure 4 is the root mean square error RMSE and determination coefficient of different CNN-Transformer models and the score function ( ), where 3×3+3×3+3×3 means using one convolution layer, and the convolution kernels in the three channels of the convolution layer are all 3×3; 1×3+1×3+1×3 means using one convolution layer, and the convolution kernels in the three channels of the convolution layer are all 1×3; 3×1+3×1+3×1 means using one convolution layer, and the convolution kernels in the three channels of the convolution layer are all 3×1; (1×2+3×1+3×3)×3 are the three convolution layers in this application.
[0037] Figure 5 This is a relationship diagram between the predicted lifetime and the actual lifetime of a terahertz photoconductive emission antenna according to a method for predicting the remaining lifetime of the terahertz photoconductive emission antenna in Example 1. DETAILED DESCRIPTION
[0038] The present invention will be further described below through specific examples.
[0039] Example 1
[0040] like Figure 1 As shown, a method for predicting the remaining lifetime of a terahertz photoconductive emission antenna includes the following steps.
[0041] (1) Data acquisition: As shown in Table 1, the terahertz time-domain spectroscopy system is turned on to collect the temperature of the femtosecond laser semiconductor saturable absorber (SESAM), the temperature of the femtosecond laser chamber, the repetition rate difference of the femtosecond laser, the output power of the femtosecond laser, the temperature of femtosecond laser pumps 1-3, the current of femtosecond laser pumps 1-3, the bias source voltage, the working environment temperature of the terahertz time-domain spectroscopy system, the working environment humidity of the terahertz time-domain spectroscopy system, and the vibration limit state data of the terahertz time-domain spectroscopy system as sensor data.
[0042] Table 1
[0043] In this embodiment, according to step (1), sensor data of three normal transmitting antennas and three damaged transmitting antennas are collected when they are continuously working under different working conditions. The working conditions are divided according to the state of the delay line. In this embodiment, the states of the delay line are 60 Hz / 20 ps, 40 Hz / 40 ps, 30 Hz / 90 ps, 25 Hz / 120 ps, 10 Hz / 120 ps and 10 Hz / 90 ps.
[0044] Assume that the timing data of the antenna operation is , the life cycle is , the sensor data dimension is (In this embodiment ), then the corresponding period of Dimensional data is , then the timing data of any transmitting antenna under a certain working condition can be expressed as X.
[0045]
[0046] (2) Data processing: First, the K-means algorithm is used to process the sensor data. Clustering is performed, and then normalization is performed based on min-max. The time sliding window is used to divide the data set into multiple time windows to obtain data set samples under different working conditions. The data set is divided into training set, validation set and test set. Finally, each sample is RUL labeled.
[0047] Step (2) specifically includes the following steps: (201) Considering the complexity of the sensor data of the terahertz time-domain spectroscopy system under different working conditions, in order to eliminate the influence of the operating conditions on the original data sequence, the sensor data is firstly processed by the K-means algorithm. Clustering is performed and then normalization is performed based on min-max, as shown in formula (2).
[0048]
[0049] in, 1 to A positive integer between 1 to A positive integer between and They are The maximum and minimum values in .
[0050] After normalization, the sensor data X is obtained , for.
[0051]
[0052] (202) Figure 3 As shown, a time sliding window is used to Processing is performed to reflect the attenuation time relationship of the transmitting antenna of different sensing signals in the process of terahertz wave generation. The sliding window length is set to , sliding step length , the total sampling length is , the data Split into A time window.
[0053] As shown in Table 2, after the above processing, 20,000 data sets were obtained for the antenna under each working condition, including 10,000 normal working data (data corresponding to the continuous operation of a normal transmitting antenna) and 10,000 abnormal working data (data corresponding to the continuous operation of a damaged normal transmitting antenna). The data were divided into training set, test set, and validation set according to the grouping ratio of 6:2:2, totaling 120,000 data under the six working conditions.
[0054] Table 2
[0055] (203) A segmented method is used to mark the lifetime of the transmitting antenna samples, as shown in formula (4). Under normal circumstances, The antenna will not experience signal attenuation or the signal has very little attenuation within the time range. After a period of time, the antenna will experience significant signal attenuation. Based on the current experience of using the transmitting antenna, set It is 700h.
[0056]
[0057] (3) The training set data is input into the CNN-Transformer model for model training, the model parameters are adjusted through the validation set, and the test set data is input into the trained CNN-Transformer model for result prediction. During the process, the asymmetric convolutional neural network is used to extract features from the sensor data after data processing in step (2), and the Transformer algorithm is used to perform time series prediction on the extracted features.
[0058] Step (3) specifically includes the following steps: (301) the asymmetric convolutional neural network includes the first convolutional layer, the second convolutional layer and the third convolutional layer, the first channel of the first convolutional layer uses a 1×3 convolution kernel to extract the relevant features between different time nodes in the time window, the second channel uses a 3×1 convolution kernel to extract the features between the sensor data in the time window, and the third channel uses a 3×3 convolution kernel to extract the data relationship between different time nodes and adjacent data in the time window, and then uses formula (5) to superimpose the results of the 1×3 convolution kernel, the 3×1 convolution kernel and the 3×3 convolution kernel. In the process, the ReLU self-normalization activation function and the batch normalization method are used.
[0059]
[0060] in, is the feature map after superposition of the first convolutional layer, They correspond to the asymmetric convolution feature maps of channels 1 to 3 respectively, and c is the number of corresponding channels.
[0061] (302) The second convolutional layer processes the output of the first convolutional layer in the same way as in step (301) to produce the output feature map. The third convolutional layer uses the same method as step (301) to process the output of the second convolutional layer and output the feature map , and finally connect the feature maps of the three convolutional layers, as shown in formula (6).
[0062]
[0063] in, is the feature map output after connection.
[0064] (303) Feature map First, it passes through the Flatten layer and is converted into a one-dimensional feature vector, and then it is connected to the Liner layer, and the feature vector is fixed to Dimension, and then position encoding. Position encoding is used to add the location information of the data, and the encoding method uses sine and cosine functions, such as formula (7) (8):
[0065]
[0066]
[0067] Where, is the time step of the window (i.e., sliding step S), is any vector in the time step, is the dimension of the feature vector.
[0068] (304) Position coding will be added The dimensional feature vector is input as the feature vector Y into the encoder layer. The encoder layer consists of a multi-head attention mechanism, a first-layer normalization, a first residual connection, a feedforward neural network, a second-layer normalization, a second residual connection, and a linear regression layer. In the multi-head attention mechanism, the input feature vector is linearly transformed to generate a query vector Q, a key vector K, and a value vector V. The attention matrix is obtained by the scaled dot product of the query vector Q and the key vector K. The matrix is multiplied with the value vector V to obtain weighted data. The process is as shown in formula (9):
[0069]
[0070] The multi-head attention mechanism consists of n single attention mechanisms, each attention head calculates the score independently. Attention mechanism output for:
[0071] (10).
[0072] Finally, all the outputs of the multi-head attention mechanism are concatenated and linearly transformed to obtain the fused feature representation (11).
[0073] Where, represents the dimension of the key vector K, 、 、 is the weight matrix in the training process, which can be continuously learned as training progresses. is a positive integer from 1 to n, is the weight matrix of multi-head attention, For the The output of the attention mechanism, It is the output after the fusion of the multi-head attention mechanism.
[0074] The first layer of the multi-head attention mechanism's output and input are normalized. Layer normalization is similar to batch normalization, but differs in that layer normalization calculates the mean and variance of the same sample across all time steps, while batch normalization calculates the mean and variance of different samples at the same time step. Compared to batch normalization, layer normalization is more adaptable to the distribution characteristics of sequential data.
[0075] (305) The first layer normalized output is fed forward to the neural network output via the first residual connection. The first residual connection is a skip connection that directly passes the input to the back of the module to avoid the gradient vanishing problem during training.
[0076] (306) The feedforward neural network is used to map the output of the multi-head attention mechanism and the first residual connection to a high-dimensional space, and then map more detailed information in the high-dimensional space to a low-dimensional space through a linear layer to explore deeper features. Its calculation method is as shown in Equation (12).
[0077]
[0078] in, and are the weight matrices of the two linear layers of the feedforward neural network, and are the bias vectors of the two linear layers of the feedforward neural network, is the input of the feedforward neural network.
[0079] The output of the feedforward neural network is normalized by the second layer and then input to the linear regression layer through the second residual connection. The second residual connection passes the input directly to the back of the module through the jump connection to avoid the gradient disappearance problem. The linear regression layer maps the knowledge learned by the encoder layer to the predicted value of RUL , calculated as in formula (13).
[0080]
[0081] in is the input of the linear regression layer, is the weight matrix, is the bias vector.
[0082] (307) Using the root mean square error (RMSE), determination coefficient ( ) and the score function ( ) as the basic evaluation criteria, such as formula (14)-(16). Among them, RMSE can effectively reflect the prediction accuracy of the algorithm. It can reflect the overall consistency between the model prediction results and the actual measured values. The closer it is to 1, the stronger the ability of the independent variable to explain the dependent variable. It can reflect whether the current prediction result is within the acceptable prediction range. The smaller the value, the more accurate the prediction result.
[0083]
[0084]
[0085]
[0086] Where N represents the number of test samples, The test sample is The predicted value of RUL when ), The test sample is The actual RUL value is is the predicted mean value of RUL.
[0087] Based on the magnitude of the prediction error, predictions can be categorized as early predictions, delayed predictions, and accurate predictions. Based on the above analysis, we can see that the proposed model can achieve more accurate predictions and avoid the problem of the model predicting a lifespan longer than the actual lifespan, preventing the device from continuing to operate after the antenna has reached the end of its lifespan, causing unnecessary usage issues.
[0088] Figure 4 To obtain the root mean square error (RMSE) and coefficient of determination ( ) score function( ).
Claims
1. A method for predicting the remaining lifetime of a terahertz photoconductive transmitting antenna, characterized in that: The following steps are included: (1) Data acquisition: Turn on the terahertz time-domain spectroscopy system and collect sensor data when the transmitting antenna works continuously under different working conditions. The working conditions are divided according to the state of the delay line. The sensor data include the temperature of the semiconductor saturable absorber of the femtosecond laser, the temperature of the femtosecond laser chamber, the repetition rate difference of the femtosecond laser, the output power of the femtosecond laser, the temperature of the femtosecond laser pumps 1-3, the current of the femtosecond laser pumps 1-3, the bias source voltage, the working environment temperature of the terahertz time-domain spectroscopy system, the working environment humidity of the terahertz time-domain spectroscopy system, and the vibration limit state data of the terahertz time-domain spectroscopy system. The time series sensor data under a certain working condition of any transmitting antenna is expressed as X: ； (2) Data processing: First, the K-means algorithm is used to process the sensor data. Perform clustering and then normalize based on min-max. Use a time sliding window to divide the data into multiple time windows to obtain data set samples under different working conditions. The data set is divided into training set, validation set and test set. Finally, each sample is RUL labeled. (3) The training set data is input into the CNN-Transformer model for model training, the model parameters are adjusted through the validation set, and the test set data is input into the trained CNN-Transformer model for result prediction. During the process, the asymmetric convolutional neural network is used to extract features from the sensor data after data processing in step (2), and the Transformer algorithm is used to perform time series prediction on the extracted features.
2. The method for predicting the remaining lifetime of a terahertz photoconductive transmitting antenna according to claim 1, wherein: Step (3) specifically includes the following steps: (301) the asymmetric convolutional neural network includes a first convolutional layer, a second convolutional layer and a third convolutional layer, the first channel of the first convolutional layer uses a 1×3 convolution kernel to extract the relevant features between different time nodes in the time window, the second channel uses a 3×1 convolution kernel to extract the features between the sensor data in the time window, and the third channel uses a 3×3 convolution kernel to extract the data relationship between different time nodes and adjacent data in the time window, and then uses formula (5) to superimpose the results of the 1×3 convolution kernel, the 3×1 convolution kernel and the 3×3 convolution kernel, and uses the ReLU self-normalization activation function and the batch normalization method in the process; ； in, is the feature map after superposition of the first convolutional layer, The asymmetric convolution feature maps corresponding to channels 1 to 3, respectively, c is the number of corresponding channels; (302) The second convolutional layer processes the output of the first convolutional layer in the same way as in step (301) to produce the output feature map. The third convolutional layer uses the same method as step (301) to process the output of the second convolutional layer and output the feature map , and finally connect the feature maps of the three convolutional layers, as shown in formula (6): ； in, is the feature map output after connection; (303) Feature map First, it passes through the Flatten layer and is converted into a one-dimensional feature vector, and then it is connected to the Liner layer, and the feature vector is fixed to Dimension, and then position coding is performed; position coding is used to add the location information of the data, and the coding method uses sine and cosine functions, such as formulas (7) and (8): ； ； Where, is the time step of the window, is any vector in the time step, is the dimension of the feature vector; (304) Position coding will be added The dimensional feature vector is used as the input feature vector Y and input into the encoder layer. The encoder layer consists of a multi-head attention mechanism, a first layer normalization, a first residual connection, a feedforward neural network, a second layer normalization, a second residual connection and a linear regression layer. The encoder layer learns the knowledge mapping to the predicted value of RUL , calculated as formula (13): ； in, is the input of the linear regression layer, is the weight matrix, is the bias vector.
3. The method for predicting the remaining lifetime of a terahertz photoconductive transmitting antenna according to claim 1, wherein: Step (2) specifically includes the following steps: (201) Considering the complexity of the sensor data of the terahertz time-domain spectroscopy system under different working conditions, in order to eliminate the influence of the operating conditions on the original data sequence, the sensor data is firstly processed by the K-means algorithm. Clustering is performed, and then normalization is performed based on min-max, as shown in formula (2). ； in, 1 to A positive integer between 1 to A positive integer between and They are The maximum and minimum values in ; After normalization, the sensor data X is obtained , for: ； (202) Using a time sliding window Processing is performed to reflect the attenuation time relationship of the transmitting antenna of different sensing signals in the process of terahertz wave generation. The sliding window length is set to , sliding step length , the total sampling length is , the data Split into time windows; (203) A segmented method is used to mark the lifetime of the transmitting antenna samples, as shown in formula (4), setting 700h; 。