Fault Diagnosis Method for Electronic DC Loads Based on Multi-Expert Distillation Network
Through multi-expert distillation network and expert variational modal decomposition algorithm, combined with global and regional attention networks, the rapid and accurate diagnosis of electronic DC load failures is achieved, and the problems of low efficiency and poor accuracy of traditional methods are solved, providing a high efficiency fault diagnosis system.
Patent Information
- Application Number
- CN202411187340.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-28
- Publication Date
- 2024-12-13
- Estimated Expiration
- 2044-08-28
AI Technical Summary
Traditional electronic DC load fault diagnosis methods are inefficient and rely on manual labor, resulting in slow and inaccurate diagnosis speed, affecting test accuracy and equipment safety.
A multi-expert distillation network is adopted to decompose current data through an expert variational modal decomposition algorithm, combine global and regional expert attention networks to extract features, and supervised learning through distillation function, input a dual feature classification network to achieve fault diagnosis.
It improves the accuracy and stability of electronic DC load fault diagnosis, provides a fast and effective fault diagnosis tool, and enhances diagnostic efficiency and system robustness.
Smart Images
Figure CN118707303B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of artificial intelligence fault diagnosis of electronic DC loads, and in particular relates to an electronic DC load fault diagnosis method based on a multi-expert distillation network. Background Art
[0002] In modern industrial and electronic systems, electronic DC loads are widely used in power supply testing, equipment debugging, and system verification. They play a vital role in ensuring the reliability and performance of equipment and systems. However, in the long-term operation process, electronic DC loads may fail due to various reasons, such as open circuit faults, short circuit faults, and overload faults. These faults not only affect the accuracy of testing and verification, but may also cause equipment damage or system downtime, resulting in serious economic losses and safety hazards. Therefore, it is particularly important to perform effective fault diagnosis on electronic DC loads.
[0003] In recent years, with the rapid development of electronic technology and computing power, fault diagnosis technology has also made significant progress. Traditional fault diagnosis methods mainly rely on the observation and analysis of fault phenomena by experienced engineers, but this method has disadvantages such as low efficiency, reliance on manual labor, and slow diagnosis speed. In order to improve the efficiency and accuracy of fault diagnosis, modern fault diagnosis technology has gradually introduced a variety of advanced methods and means. The use of deep learning networks can improve the accuracy of fault diagnosis of electronic DC loads. At the same time, the use of multi-expert distillation networks in deep learning networks for fault diagnosis can greatly improve the efficiency and stability of diagnosis. Summary of the invention
[0004] The purpose of the present invention is to provide a fault diagnosis method for an electronic DC load. The method adopts a multi-expert distillation deep learning network, which can quickly, accurately and stably diagnose the fault type of the electronic DC load, and provides an effective tool for the electronic DC load fault diagnosis system.
[0005] In order to solve the above technical problems, the present invention provides an electronic DC load fault diagnosis method, comprising:
[0006] Collect the current data of the electronic DC load circuit, use the expert variational mode decomposition algorithm to decompose the current data into K IMF signals, and organize them into an electronic DC load diagnostic data set;
[0007] A global expert attention network is used to extract global features of the electronic DC load diagnosis dataset;
[0008] A regional expert attention network is used to extract regional features of the electronic DC load diagnosis dataset;
[0009] The distillation function is used to supervise the mutual learning between the global expert attention network and the regional expert attention network, and the global enhanced features and regional enhanced features of the electronic DC load diagnosis dataset are obtained through the learned network;
[0010] The global enhanced features and regional enhanced features of the electronic DC load diagnosis dataset are input into the dual-feature classification network to obtain the final electronic DC load fault diagnosis results.
[0011] Optionally, the process of collecting current data of the electronic DC load circuit, decomposing the current data into K IMF signals by using an expert variational mode decomposition algorithm, and arranging to form an electronic DC load diagnostic data set includes:
[0012] The current data of the electronic DC load circuit during operation is collected, and the expert variational mode decomposition algorithm is used for the current data to establish an expert variational model. The penalty factor and multiplication operator are used to transform the model solution process into an unconstrained optimization problem. The alternating direction multiplier method (ADMM) is used to solve the optimal solution of the model problem. Through iterative solution process, K IMF signals of current data decomposition are finally obtained. The electronic DC load diagnostic data set is formed according to the three fault types of open circuit fault, short circuit fault and overload fault.
[0013] Optionally, a global feature extraction process of the electronic DC load diagnosis dataset using a global expert attention network includes:
[0014] The samples of the electronic DC load diagnosis dataset are input into the global expert attention network. The global expert attention network includes a convolutional feature layer, a dual-channel attention layer and a fusion layer. The convolutional feature layer simultaneously extracts the window features of K IMF signals of the diagnosis samples through a sliding convolution window. The key, value and query of the dual channels perform weighted summation on the window features of K IMF signals of the diagnosis samples to obtain two channel features. The fusion layer performs nonlinear mapping and fusion of the two channel features through full connection to obtain the global features of the electronic DC load diagnosis dataset samples.
[0015] Optionally, a regional expert attention network is used to extract regional features of the electronic DC load diagnosis dataset, including:
[0016] The samples of the electronic DC load diagnosis dataset are input into the regional expert attention network. The regional expert attention network contains a regional attention layer and a transformation matrix. The regional attention layer contains K attention blocks with the same structure, where each attention block extracts the component features corresponding to one IMF signal. The regional attention layer extracts K component features corresponding to the K IMF signals in the sample. The K component features are simultaneously multiplied by the transformation matrix to obtain the regional features of the electronic DC load diagnosis dataset samples.
[0017] Optionally, a distillation function is used to supervise the global expert attention network and the regional expert attention network to learn from each other, and the global enhanced features and regional enhanced features of the electronic DC load diagnosis dataset are obtained through the learned network, including:
[0018] The distillation function is used to calculate the feature distance between global features and regional features. The global expert attention network and the regional expert attention network continuously update their respective network parameters according to the back-propagated feature distance to achieve mutual learning between the global expert attention network and the regional expert attention network. The electronic DC load diagnosis dataset is re-input into the learned global expert attention network and the regional expert attention network to obtain the global enhanced features and regional enhanced features of the electronic DC load diagnosis dataset samples.
[0019] Optionally, the global enhanced features and the regional enhanced features of the electronic DC load diagnosis data set are input into a dual-feature classification network to obtain a final electronic DC load fault diagnosis result, including:
[0020] The global enhanced features and regional enhanced features of the electronic DC load diagnosis dataset samples are input into a dual-feature classification network. The dual-feature classification network contains a convolutional diagnosis layer and a fully connected diagnosis layer running in parallel. The convolutional diagnosis layer classifies the input global features and outputs the convolution classification probability. The fully connected diagnosis layer classifies the input local features and outputs the fully connected classification probability. The dual-feature classification network finally performs a weighted summation of the convolution classification probability and the fully connected classification probability to obtain the fault diagnosis result of the electronic DC load sample.
[0021] The present application discloses an electronic DC load fault diagnosis method based on a multi-expert distillation network, including using an expert variational mode decomposition algorithm to decompose current data into K IMF signals, and organize them to form an electronic DC load diagnostic data set; inputting diagnostic data set samples into a global expert attention network to extract global features of samples; using a regional expert attention network to extract regional features of diagnostic data set samples; using a distillation function to supervise the mutual learning of global expert attention and regional expert attention networks, inputting diagnostic samples into the two learned networks again, and obtaining global enhanced features and regional enhanced features of diagnostic data set samples; inputting global enhanced features and regional enhanced features into a dual feature classification network to obtain the final electronic DC load fault diagnosis result. The electronic DC load fault diagnosis method of the present application has high accuracy and strong robustness, and provides an effective tool for the electronic DC load intelligent fault diagnosis system. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0023] Figure 1 A schematic flow chart of the electronic DC load fault diagnosis method provided by the present invention;
[0024] Figure 2 A global expert attention network framework diagram of the electronic DC load fault diagnosis method provided by the present invention;
[0025] Figure 3 A regional expert attention network framework diagram of the electronic DC load fault diagnosis method provided by the present invention;
[0026] Figure 4 This is a network flow chart of the electronic DC load fault diagnosis method provided by the present invention. DETAILED DESCRIPTION
[0027] In order to enable those skilled in the art to better understand the scheme of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0028] like Figure 1 As shown, Figure 1 A flow chart of a method for extracting features of motor uneven breath fault provided by the present invention, the method comprising:
[0029] S11: Collect current data of the electronic DC load circuit, use the expert variational mode decomposition algorithm to decompose the current data into K IMF signals, and organize them into an electronic DC load diagnostic data set.
[0030] It should be noted that the present application introduces the expert variational mode decomposition algorithm to decompose the current data into K intrinsic mode component (IMF) signals because the decomposed K IMF signals contain more stable, easier to identify and more comprehensive electronic DC load fault characteristics, thereby ensuring the robustness and accuracy of the subsequent feature extraction network.
[0031] The current sensor is installed in the running electronic DC load circuit to collect the circuit current signal with a sampling frequency of N kHz.
[0032] The expert variational mode decomposition algorithm is used to decompose the current data into K IMF signals, which are then organized into an electronic DC load diagnostic data set. This includes five steps:
[0033] Step 1: In order to decompose the intrinsic modal components representing the fault characteristics of the electronic DC load from the original current data, the present application adopts the expert variational model shown in equations (1, 2):
[0034]
[0035]
[0036] In the formula, is the partial derivative operation, is the Dirichlet distribution function, ∗ represents convolution, represents the K IMFs obtained by the expert variational mode decomposition algorithm, represents the center frequency of each IMF, f represents the original current signal, and K is the number of IMFs.
[0037] Step 2: Use the penalty factor and multiplication operator to transform the solution process of the model in equation (1) into an unconstrained optimization problem, and obtain the augmented Lagrangian function as shown in equation (3):
[0038]
[0039] In the formula, is the Lagrange constant, is the penalty factor.
[0040] Step 3: In order to solve the optimal solution of the expert variational model problem, this application uses the alternating direction multiplier method (ADMM) to update the variables and establish the solution process shown in formula (4):
[0041]
[0042] In the formula is the noise tolerance, and Corresponding to The Fourier transform of .
[0043] Step 4: Initialization , iterate by n = n+1, where n is the number of iterations, and update the parameters Repeat this step until the iteration stop condition is met and the optimal , where is a constant and ;
[0044]
[0045] Step 5: Finally, we get K IMF signals of current data decomposition ,According to the three fault types of open circuit fault, short circuit fault and overload fault, the electronic DC load diagnosis data set is formed.
[0046] Based on the above discussion, in an optional embodiment of the present application, the process of setting the current data into K IMFs using the expert variational mode decomposition algorithm may include:
[0047] In an optional embodiment of S11, the current sampling frequency is 15 kHz, the current data is decomposed into 9 IMF signals, the number of iterations n is 120, and the iteration stop condition ε=0.5.
[0048] S12: Global expert attention network is used to extract global features of electronic DC load diagnosis dataset.
[0049] It should be noted that this application introduces a global expert attention network to extract global features of samples because the global expert attention network has a global vision and the attention mechanism in the network can allocate more computing resources to important feature information.
[0050] The global expert attention network structure is as follows Figure 2 As shown, the network includes a convolutional feature layer, a dual-channel attention layer and a fusion layer. The network extracts the global features of the electronic DC load diagnosis data set samples, which specifically includes three steps:
[0051] Step 1: The electronic DC load diagnosis dataset samples are first input into the convolutional feature layer of the global expert attention network. The convolutional feature layer simultaneously performs sliding convolution on the K IMF signals of the samples to obtain window features.
[0052] Step 2: The window features are input into the dual-channel attention layer of the global expert attention network. The parameters of the dual channels in the dual-channel attention layer are independent of each other. The key, value and query of the dual channels are weighted summed to obtain two channel features.
[0053] Step 3: Input the two channel features into the fusion layer of the network. The fusion layer performs nonlinear mapping fusion on the dual channel features through full connection to obtain the global features of the electronic DC load diagnosis data set samples.
[0054] Based on the above discussion, in an optional embodiment of the present application, the global feature setting process for extracting the electronic DC load diagnostic data set samples may include:
[0055] In an optional embodiment of S12, the Adam optimization algorithm is used for training, the learning rate is 0.001, the input epoch is 64, and the number of iterations is 1000.
[0056] S13: Regional Expert Attention Network is used to extract regional features of the electronic DC load diagnosis dataset.
[0057] It should be noted that this application introduces the regional expert attention network to extract regional features of samples because the regional expert attention network has a more accurate and delicate regional vision than the global expert attention network, and can extract more detailed regional features from the samples.
[0058] The regional expert attention network structure is as follows Figure 3 As shown, the regional expert attention network includes a regional attention layer and a transformation matrix. The regional expert attention network extracts regional features of samples of the electronic DC load diagnosis dataset, which specifically includes two steps.
[0059] Step 1: The electronic DC load diagnosis dataset samples first enter the regional attention layer of the regional expert attention network. The regional attention layer contains K attention blocks with the same structure, where each attention block extracts the component features corresponding to one IMF signal. The regional attention layer extracts K component features corresponding to the K IMF signals in the sample.
[0060] Step 2: After transforming the matrix dot multiplication of the K component features of the sample, the K component features of the sample of the electronic DC load diagnosis dataset are linearly transformed together into the regional features of the sample.
[0061] Based on the above discussion, in an optional embodiment of the present application, the process of setting the regional features for extracting the electronic DC load diagnostic data set samples may include:
[0062] In an optional embodiment of S13, the Adam optimization algorithm is used for training, the learning rate is 0.003, the input epoch is 64, and the number of iterations is 500.
[0063] S14: The distillation function is used to supervise the global expert attention network and the regional expert attention network to learn from each other, and the global enhanced features and regional enhanced features of the electronic DC load diagnosis dataset are obtained through the learned network.
[0064] It should be noted that the present application introduces the distillation function to calculate the distance between global features and regional features because the global expert attention network and the regional expert attention network learn from each other through distillation loss to make up for their respective shortcomings, thereby obtaining a global expert attention network and a regional expert attention network with high-accuracy feature extraction capabilities.
[0065] The distillation function is used to calculate the distance between the global features and regional features of the sample, and the global expert attention network and the regional expert attention network are supervised to learn from each other, and finally the global enhanced features and regional enhanced features of the diagnosis sample are obtained. Specifically, there are four steps:
[0066] Step 1: In order to calculate the distance between the global features and the regional features of the diagnostic sample, this application establishes the distillation function shown in formula (6):
[0067]
[0068] In the formula , are the global and regional features of the diagnostic samples, respectively. represents the JS divergence, represents the mean square error.
[0069] Step 2: In order to convert the distance between the global features and regional features of the diagnostic sample into the optimization parameters of the global expert attention network and the regional expert attention network, it is necessary to calculate The gradient of is calculated as shown in formula (7):
[0070]
[0071] In the formula represents the derivative, and W represents the network parameters.
[0072] Step 3: The global expert attention network and the regional expert attention network continuously update the network parameters according to the gradient calculated by formula (7) until the iteration stops. The network parameter update formula is shown in formula (8):
[0073]
[0074] In the formula Represents the learning rate.
[0075] Step 4: Input the electronic DC load diagnosis dataset into the global expert attention network and the regional expert attention network after updating the network parameters to obtain the global enhanced features and regional enhanced features of the diagnosis samples.
[0076] Based on the above discussion, in an optional embodiment of the present application, for the above-mentioned use of the distillation function to calculate the distance between the global features and the regional features of the sample, the global expert attention network and the regional expert attention network are supervised to learn from each other, and the setting process of finally obtaining the global enhanced features and the regional enhanced features of the diagnosis sample may include:
[0077] In an optional embodiment of S14, the Adam optimization algorithm is used for training, the learning rate is 0.001, the input epoch is 64, and the number of iterations is 1000.
[0078] S15: The global enhanced features and regional enhanced features of the electronic DC load diagnosis dataset are input into the dual-feature classification network to obtain the final electronic DC load fault diagnosis results.
[0079] It should be noted that the dual-feature classification network introduced in this application can perform high-accuracy and high-robust fault diagnosis on electronic DC load circuits because the dual-feature classification network has a convolutional diagnostic layer and a fully connected diagnostic layer, which can fully utilize the global enhanced feature and local enhanced feature information in the diagnostic samples at the same time.
[0080] A dual-feature classification network is used to perform fault diagnosis on electronic DC load circuits based on the global enhanced features and regional enhanced features of the diagnosis samples. The specific steps include:
[0081] Step 1: Establish a dual-feature classification network. The dual-feature network contains a convolutional diagnosis layer and a fully connected diagnosis layer running in parallel. The global enhanced features and regional enhanced features of the diagnostic samples are input into the dual-feature classification network. The convolutional diagnosis layer in the dual-feature classification network performs multi-layer convolution on the global features to extract convolutional features, and the convolutional features are input into the softmax classifier to obtain the convolutional softmax classification probability.
[0082] Step 2: The fully connected diagnostic layer in the dual-feature classification network performs nonlinear mapping on the regional enhanced features of the diagnostic samples to extract the fully connected features, and inputs the fully connected features into the softmax classifier to obtain the fully connected softmax classification probability.
[0083] Step 3: The dual-feature classification network performs weighted summation on the convolution classification probability and the fully connected classification probability obtained in the above steps to obtain the fault diagnosis result of the electronic DC load sample.
[0084] Based on the above discussion, in an optional embodiment of the present application, for the above-mentioned use of the dual-feature classification network to perform fault diagnosis on the electronic DC load circuit based on the global enhanced features and regional enhanced features of the diagnostic samples, the setting process may include:
[0085] In an optional embodiment of S15, the Adam optimization algorithm is used for training, the learning rate is 0.001, the input epoch is 32, the number of iterations is 100, and the weights of the convolution classification probability and the fully connected classification probability are 0.3 and 0.7 respectively.
[0086] Figure 4The network flow chart of the electronic DC load fault diagnosis method provided by the present application, the present application first uses the expert variational mode decomposition algorithm to decompose the current data into K intrinsic mode components (IMF), and organizes them into an electronic DC load diagnosis data set; then, the electronic DC load diagnosis data set samples are input into the global expert attention network to extract the global features of the samples; then the regional expert attention network is used to extract the regional features of the samples; the distance between the global features and the regional features of the diagnosis samples is calculated by the distillation function, and the distance is used to supervise the mutual learning of the global expert attention network and the regional expert attention network, and the diagnosis samples are input into the learned global expert attention network and the regional expert attention network again to obtain the global enhanced features and regional enhanced features of the diagnosis samples; finally, the global enhanced features and regional enhanced features of the diagnosis samples are input into the dual feature classification network to obtain the final electronic DC load fault diagnosis results. The electronic DC load fault diagnosis method provided by the present application can quickly, accurately and stably diagnose the fault type of the electronic DC load, and provides an effective tool for the electronic DC load fault diagnosis system.
[0087] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Specific examples are used herein to illustrate the principles and implementation methods of the present invention, and the description of the above embodiments is only used to help understand the method of the present invention and its core idea. It should be pointed out that for ordinary technicians in this technical field, without departing from the principle of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.
Claims
1. An electronic DC load fault diagnosis method based on a multi-expert distillation network is characterized in that: include: Collect the current data of the electronic DC load circuit, use the expert variational mode decomposition algorithm to decompose the current data into K IMF signals, and organize the decomposed IMF signals according to open circuit fault, short circuit fault and overload fault to form an electronic DC load diagnosis data set; The global expert attention network extracts the global features of the IMF signal of the electronic DC load diagnosis dataset through the convolutional feature layer, the dual-channel attention layer and the fusion layer; The regional expert attention network extracts the component features of the IMF signal of the electronic DC load diagnosis dataset through the attention block, and linearly maps the component features to regional features through the transformation matrix; The distillation function is used to supervise the mutual learning between the global expert attention network and the regional expert attention network, and the global enhanced features and regional enhanced features of the electronic DC load diagnosis dataset are obtained through the learned network; The global enhanced features and regional enhanced features of the electronic DC load diagnosis dataset are input into the dual-feature classification network to obtain the final electronic DC load fault diagnosis results.
2. The electronic DC load fault diagnosis method as claimed in claim 1, characterized in that: The current data of the electronic DC load circuit is collected, and the expert variational mode decomposition algorithm is used to decompose the current data into K IMF signals. The decomposed IMF signals are sorted according to open circuit faults, short circuit faults and overload faults to form an electronic DC load diagnosis data set, including: The current data of the electronic DC load circuit during operation is collected, and the expert variational mode decomposition algorithm is used for the current data to establish an expert variational model. The penalty factor and multiplication operator are used to transform the model solution process into an unconstrained optimization problem. The alternating direction multiplier method is used to solve the optimal solution of the model problem. Through iterative solution process, K IMF signals of current data decomposition are finally obtained. The electronic DC load diagnostic data set is formed according to the three fault types of open circuit fault, short circuit fault and overload fault.
3. The electronic DC load fault diagnosis method as claimed in claim 1, characterized in that: The global expert attention network extracts the global features of the IMF signal of the electronic DC load diagnosis dataset through the convolutional feature layer, the dual-channel attention layer and the fusion layer, including: The samples of the electronic DC load diagnosis dataset are input into the global expert attention network. The global expert attention network includes a convolutional feature layer, a dual-channel attention layer and a fusion layer. The convolutional feature layer simultaneously extracts the window features of K IMF signals of the diagnosis samples through a sliding convolution window. The key, value and query of the dual channels perform weighted summation on the window features of K IMF signals of the diagnosis samples to obtain two channel features. The fusion layer performs nonlinear mapping and fusion of the two channel features through full connection to obtain the global features of the electronic DC load diagnosis dataset samples.
4. The electronic DC load fault diagnosis method as claimed in claim 1, characterized in that: The regional expert attention network extracts the component features of the IMF signal of the electronic DC load diagnosis dataset through the attention block, and linearly maps the component features to regional features through the transformation matrix, including: The samples of the electronic DC load diagnosis dataset are input into the regional expert attention network. The regional expert attention network contains a regional attention layer and a transformation matrix. The regional attention layer contains K attention blocks with the same structure, where each attention block extracts the component features corresponding to one IMF signal. The regional attention layer extracts K component features corresponding to the K IMF signals in the sample. The K component features are simultaneously multiplied by the transformation matrix to obtain the regional features of the electronic DC load diagnosis dataset samples.
5. The electronic DC load fault diagnosis method as claimed in claim 1, characterized in that: The distillation function is used to supervise the mutual learning of the global expert attention network and the regional expert attention network, and the global enhanced features and regional enhanced features of the electronic DC load diagnosis dataset are obtained through the learned network, including: The distillation function is used to calculate the feature distance between global features and regional features. The global expert attention network and the regional expert attention network continuously update their respective network parameters according to the back-propagated feature distance to achieve mutual learning between the global expert attention network and the regional expert attention network. The electronic DC load diagnosis dataset is re-input into the learned global expert attention network and the regional expert attention network to obtain the global enhanced features and regional enhanced features of the electronic DC load diagnosis dataset samples.
6. The electronic DC load fault diagnosis method as claimed in claim 1, characterized in that: The global enhanced features and regional enhanced features of the electronic DC load diagnosis dataset are input into the dual-feature classification network to obtain the final electronic DC load fault diagnosis results, including: The global enhanced features and regional enhanced features of the electronic DC load diagnosis dataset samples are input into a dual-feature classification network. The dual-feature classification network contains a convolutional diagnosis layer and a fully connected diagnosis layer running in parallel. The convolutional diagnosis layer classifies the input global features and outputs the convolution classification probability. The fully connected diagnosis layer classifies the input local features and outputs the fully connected classification probability. The dual-feature classification network finally performs a weighted summation of the convolution classification probability and the fully connected classification probability to obtain the fault diagnosis result of the electronic DC load sample.