Feature extraction method of synchronous waveform data, model training method, medium and equipment

Abstract

The invention discloses a synchronous waveform data feature extraction method, a model training method, a medium and equipment, and the method comprises the steps: calculating the fluctuation feature of synchronous waveform data of each sample through employing a first sliding window with a larger size, and screening out a suspected fault region, thereby positioning the fault region, and carrying out the recognition of the fault region. And the identification precision of the fault area is effectively improved. And then in the suspected fault area, using a second sliding window with a small size to extract local features, and fusing the local features of different samples in a point-by-point convolution manner to realize efficient fusion of cross-sample features. And finally, the extracted local features are converted into feature vectors through a full connection layer, and a loss value is calculated through a feature loss function. And through iterative optimization of the loss function, the model is continuously adjusted until convergence, so that high-precision fault features are finally output. According to the method, the problems of low fault feature extraction efficiency, insufficient identification precision and insufficient cross-channel information fusion of a traditional method are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent monitoring and fault diagnosis of electric power systems, and in particular to a feature extraction method, a model training method, a medium and a device for synchronous waveform data. Background Art

[0002] In the power system, synchronous waveform data refers to the original time-domain power waveform signal collected under a unified time reference. However, when processing multi-dimensional synchronous waveform data, traditional fault feature extraction methods have difficulty efficiently and accurately locating and extracting fault feature areas in a large amount of waveform data. In this case, the fixed window is prone to miss the key fault area or add redundant information, resulting in low efficiency of fault feature extraction and insufficient recognition accuracy. In addition, in the processing of multi-channel waveform data, traditional methods have difficulty in achieving effective fusion of cross-channel information, further limiting the effect of fault feature extraction. Summary of the invention

[0003] Based on this, it is necessary to provide feature extraction methods, model training methods, media and equipment for synchronized waveform data to solve the problems of low efficiency in fault feature extraction, insufficient recognition accuracy and difficulty in achieving effective fusion of cross-channel information.

[0004] A training method for an adaptive window convolutional network, the adaptive window convolutional network comprising a sliding window layer, a depthwise separable convolution and a fully connected layer connected in sequence, the sliding window layer setting a first sliding window of a first size, the depthwise separable convolution layer setting a second sliding window of a second size, the first size being larger than the second size, the method comprising:

[0005] Acquire a sample set; wherein the sample set includes a plurality of samples, each sample including a plurality of sample synchronous waveform data within a preset time period;

[0006] Input the samples into the adaptive window convolution network in sequence, traverse and calculate the fluctuation characteristic value of each sample synchronization waveform data in the window through the first sliding window of the sliding window layer, and screen the suspected fault area of ​​each sample synchronization waveform data based on the fluctuation characteristic value;

[0007] Extracting local features in the suspected fault area of ​​each sample synchronous waveform data through the second sliding window of the depthwise separable convolution layer, and performing point-by-point convolution on the local features of different sample synchronous waveform data to obtain fused features;

[0008] The fused features are converted into feature vectors of synchronized waveform data through the fully connected layer, the feature vectors of synchronized waveform data are input into a preset feature loss function to calculate a loss value, and the adaptive window convolutional network is iteratively adjusted according to the loss value until the adaptive window convolutional network converges.

[0009] In one embodiment, traversing the first sliding window of the sliding window layer to calculate the fluctuation characteristic value of each sample synchronous waveform data in the window, and screening the suspected fault area of ​​each sample synchronous waveform data based on the fluctuation characteristic value, includes:

[0010] Calculate the following when each first sliding window is sliding into the window:

[0011]

[0012] In the above formula, μ w (t) represents the mean value at time t; σ w (t) represents the standard deviation at time t; κ w (t) represents the kurtosis at time t; W1 represents the first dimension; X i (k) represents the kth data point of the i-th sample synchronous waveform data in the first sliding window;

[0013] A window whose mean is greater than a preset first threshold, and / or whose standard deviation is greater than a preset second threshold, and / or whose kurtosis is greater than a preset third threshold is regarded as a suspected fault area.

[0014] In one embodiment, the first sliding window of the sliding window layer is used to traverse and calculate the fluctuation characteristic value of each sample synchronous waveform data in the window, and the suspected fault area of ​​each sample synchronous waveform data is screened based on the fluctuation characteristic value, which is expressed as:

[0015] Calculate the following when each first sliding window is sliding into the window:

[0016]

[0017] In the above formula, γ w (t) represents the skewness at time t; R w (t) represents the extreme value difference at time t; μ XWMU represents the mean; W1 represents the first size; μ w (t) represents the mean value at time t; σ w (t) represents the standard deviation at time t; X i (k) represents the kth data point of the i-th sample synchronous waveform data in the first sliding window;

[0018] A window whose skewness is greater than a preset fourth threshold value and / or whose extreme value difference is greater than a preset fifth threshold value is taken as a suspected fault area.

[0019] In one of the embodiments, the second sliding window of the depthwise separable convolutional layer extracts local features in the suspected fault area of ​​each sample synchronous waveform data, which is expressed as:

[0020]

[0021] In the above formula, Y in represents the local features extracted from the nth suspected fault area in the i-th sample synchronous waveform data; X in (t) represents the tth data point in the nth suspected fault area in the i-th sample synchronous waveform data; W2(K) represents the convolution kernel of the second sliding window, with a size of K; X in (t+k) represents the data value at the t+kth time point in the suspected fault area; W2(k) represents the kth element of the convolution kernel.

[0022] In one embodiment, the local features of the synchronous waveform data of different samples are convolved point by point to obtain fusion features, which are expressed as:

[0023]

[0024] In the above formula, F n represents the nth fusion feature, C represents the total number of sample synchronization waveform data in each sample, and Y in represents the local features extracted from the nth suspected fault area in the i-th sample synchronous waveform data, W 1×1 Represents a point-wise convolution kernel of size 1×1.

[0025] In one embodiment, the feature loss function is expressed as:

[0026] L total =λ CE ·L CE +λ MSE ·L MSE

[0027]

[0028] In the above formula, L total represents the loss value; λ CE , MSE is a hyperparameter; L CE is the classification loss; L MSE is the regression loss; is the one-hot encoding form of the true label of the nth sample, where k is the index of the class; is the predicted probability that the nth fused feature obtained by the fully connected layer belongs to class k; N represents the total number of fused features, K is the number of categories; F′ wn represents the expected feature vector of the synchronous waveform data in the nth feature channel; F wnRepresents the feature vector extracted from the nth feature channel of the synchronous waveform data.

[0029] A method for extracting features from synchronous waveform data, the method comprising:

[0030] Acquire multiple measured synchronous waveform data within a preset time period;

[0031] The multiple measured synchronous waveform data are input into an adaptive window convolutional network to obtain feature vectors of the multiple measured synchronous waveform data; wherein the adaptive window convolutional network is trained by the above-mentioned adaptive window convolutional network training method.

[0032] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the processor executes the steps of the training method of the adaptive window convolutional network and the feature extraction method of step waveform data.

[0033] A feature extraction device for synchronous waveform data includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the training method of the above-mentioned adaptive window convolutional network and the feature extraction method of the synchronous waveform data.

[0034] The present invention provides a feature extraction method, a model training method, a medium and a device for synchronous waveform data. First, a first sliding window with a larger sliding window is used to calculate the fluctuation characteristics of each sample synchronous waveform data, and then the suspected fault area is screened out, so as to locate the fault area and effectively improve the recognition accuracy of the fault area. Then, in the suspected fault area, a second sliding window with a smaller size is used to extract local features, and the local features of different samples are fused point by point convolution to achieve efficient fusion of cross-sample features. Finally, the extracted local features are converted into feature vectors through a fully connected layer, and then the loss value is calculated through a feature loss function. Through iterative optimization of the loss function, the model is continuously adjusted until convergence, thereby finally outputting high-precision fault features. The present invention solves the problems of low efficiency in fault feature extraction, insufficient recognition accuracy, and insufficient cross-channel information fusion in traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in 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.

[0036] in:

[0037] Figure 1 A flowchart of a training method for an adaptive window convolutional network;

[0038] Figure 2 A flowchart for feature extraction of synchronous waveform data;

[0039] Figure 3 A flowchart of a feature extraction method for synchronous waveform data;

[0040] Figure 4 This is a structural block diagram of the terminal equipment. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0042] The terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally includes steps or units that are not listed, or optionally includes other steps or units inherent to these processes, methods, products or devices.

[0043] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0044] like Figure 1 As shown, Figure 1 It is a flow chart of a training method of an adaptive window convolutional network in an embodiment, wherein the adaptive window convolutional network comprises a sliding window layer, a depthwise separable convolution and a fully connected layer connected in sequence, wherein the sliding window layer sets a first sliding window of a first size, and the depthwise separable convolutional layer sets a second sliding window of a second size, wherein the first size is larger than the second size. The steps provided in the training method of the adaptive window convolutional network in this example include:

[0045] S101, obtaining a sample set.

[0046] The sample set includes multiple samples, each of which includes multiple sample synchronous waveform data within a preset time period. The sample set is the basis for model training and fault feature extraction, and provides input data to support algorithm analysis.

[0047] In this embodiment, Figure 2 As shown, the sample synchronization waveform data in each sample includes voltage and current waveforms, which are recorded as:

[0048] X(t)=[X V (t), X I (t)], t = 1, 2, ..., T

[0049] S102, input the samples into the adaptive window convolution network in sequence, traverse and calculate the fluctuation characteristic value of each sample synchronous waveform data in the window through the first sliding window of the sliding window layer, and screen the suspected fault area of ​​each sample synchronous waveform data based on the fluctuation characteristic value.

[0050] Among them, the sliding window can scan the synchronous waveform data step by step. In each scanning step, the window size and position can be adaptively adjusted according to the characteristics of the data. The first sliding window is a specific window in the sliding window layer. It slides on the sample data, traverses the time series data of each sample, and calculates the data features in each window. These features help to reveal the regularity and abnormal behavior of the data, especially when analyzing faults. Among them, the size of the first sliding window is relatively large, which is used to capture global or longer time range features. By analyzing the fluctuation feature values, the network can screen out areas that may contain fault information, that is, suspected fault areas.

[0051] Optionally, in S102, traversing the first sliding window of the sliding window layer to calculate the fluctuation characteristic value of each sample synchronous waveform data in the window, and screening the suspected fault area of ​​each sample synchronous waveform data based on the fluctuation characteristic value specifically includes the following steps:

[0052] A1. Calculate in each first sliding window:

[0053]

[0054] In the above formula, μ w (t) represents the mean value at time t, reflecting the central trend of the waveform; σ w (t) represents the standard deviation at time t, which is a measure of data volatility. The larger the value, the more drastic the waveform fluctuation. w(t) represents the kurtosis at time t, which is used to measure the sharpness of the data distribution. The larger the kurtosis, the sharper the peak of the data, which may mean that there is a sudden event in the signal; W1 represents the first size. This window size needs to be selected according to the actual situation and is usually set according to the time scale of the fault characteristics; X i (k) represents the kth data point of the i-th sample synchronization waveform data in the first sliding window. In this embodiment, i=2;

[0055] A2. The windows whose mean is greater than a preset first threshold, and / or whose standard deviation is greater than a preset second threshold, and / or whose kurtosis is greater than a preset third threshold are regarded as suspected fault areas.

[0056] These conditions indicate that the waveform at that moment may be abnormal, which may be an indication of a fault. By setting reasonable thresholds and combining multiple features for judgment, the accuracy, robustness and adaptability of fault detection can be improved, thereby helping to quickly locate and respond to fault events in the power system.

[0057] Optionally, in S102, the fluctuation characteristic value of each sample synchronous waveform data in the sliding window layer is traversed and calculated, and the suspected fault area of ​​each sample synchronous waveform data is screened based on the fluctuation characteristic value, which is expressed as:

[0058] B1. Calculate in each first sliding window:

[0059]

[0060] In the above formula, γ w (t) represents the skewness at time t, which measures the asymmetry of the data distribution. A positive value indicates that the data distribution is biased to the right, a negative value indicates that the data distribution is biased to the left, and a value close to zero indicates that the data distribution is roughly symmetrical. In fault detection, skewness can reflect whether there is obvious asymmetric fluctuation in the data; R w (t) represents the extreme value difference at time t, reflecting the amplitude of data fluctuation. The extreme value difference usually increases when a fault occurs, especially when a short circuit or overload occurs in the power system, the fluctuation amplitude often increases significantly; μ XWMU represents the mean; W1 represents the first size; μ w (t) represents the mean value at time t; σ w (t) represents the standard deviation at time t; X i (k) represents the kth data point of the i-th sample synchronization waveform data in the first sliding window;

[0061] B2. The window whose skewness is greater than a preset fourth threshold value and / or whose extreme value difference is greater than a preset fifth threshold value is regarded as a suspected fault area.

[0062] These conditions indicate that the waveform at that moment may be abnormal, which may be an indication of a fault. By setting reasonable thresholds and combining multiple features for judgment, the accuracy, robustness and adaptability of fault detection can be improved, thereby helping to quickly locate and respond to fault events in the power system.

[0063] S103, extracting local features from the suspected fault area of ​​each sample synchronous waveform data through the second sliding window of the depthwise separable convolutional layer, and performing point-by-point convolution on the local features of different sample synchronous waveform data to obtain fused features.

[0064] The depthwise separable convolution layer is designed with two operation steps. The first is to use an independent second sliding window to perform convolution on each input. The function of the second sliding window is to perform a more detailed analysis of the screened suspected fault area to extract local features. The second is that after the convolution, the point-by-point convolution uses a convolution kernel of size 1×1 to perform cross-channel mixing. This operation fuses the information of each channel to generate a new fusion feature. The fusion feature can integrate information from different channels to better represent the global characteristics of the signal and further enhance the classification or prediction ability of the model.

[0065] Optional, such as Figure 2 As shown, in s103, the second sliding window of the depth-separable convolutional layer extracts local features in the suspected fault area of ​​each sample synchronization waveform data, which is expressed as:

[0066]

[0067] In the above formula, Y in represents the local features extracted from the nth suspected fault area in the i-th sample synchronous waveform data; X in (t) represents the tth data point in the nth suspected fault area in the i-th sample synchronous waveform data; W2(K) represents the convolution kernel of the second sliding window, with a size of K; X in (t+k) represents the data value at the t+kth time point in the suspected fault area; W2(k) represents the kth element of the convolution kernel.

[0068] The convolution operation of the second sliding window can effectively extract local features in the suspected fault area and extract the local change pattern of the signal by weighted summation.

[0069] Optional, such as Figure 2 As shown, in s103, the local features of the synchronous waveform data of different samples are convolved point by point to obtain the fusion features, which are expressed as:

[0070]

[0071] In the above formula, F n represents the nth fusion feature, which contains the comprehensive information of multiple sample synchronous waveform data in the same suspected fault area; C represents the total number of sample synchronous waveform data in each sample, Y in represents the local features extracted from the nth suspected fault area in the i-th sample synchronous waveform data, W 1×1 Represents a point-by-point convolution kernel of size 1×1, that is, the convolution kernel only involves a single data point at each position without considering neighboring data points.

[0072] In this way, by performing weighted summation on the local features of different samples, the fused features obtained can integrate the key information of all samples and improve the expressiveness of the features.

[0073] S104, converting the fused features into feature vectors of the synchronized waveform data through a fully connected layer, inputting the feature vectors of the synchronized waveform data into a preset feature loss function to calculate a loss value, and iteratively adjusting the adaptive window convolutional network through the loss value until the adaptive window convolutional network converges.

[0074] In this step, the fully connected layer is used to achieve high-dimensional compression or mapping of features, thereby outputting a fixed-length feature vector for fault detection, classification, or other analysis tasks. In this scheme, the feature loss function compares the feature vector of the synchronized waveform data with the expected target feature and guides the network training process by calculating the loss value.

[0075] Optionally, the feature loss function is expressed as:

[0076] L total =λ CE ·L CE +λ MSE ·L MSE

[0077]

[0078] In the above formula, L total represents the loss value; λ CE , MSE is a hyperparameter used to adjust the relative importance of classification loss and regression loss; L CE is the classification loss, which is used to measure the category probability distribution predicted by the network and the true label distribution; L MSE is the regression loss, which is used to measure the error between the predicted value of the synchronized waveform data feature vector and the target feature vector; is the one-hot encoding form of the true label of the nth sample, where k is the index of the class; is the predicted probability that the nth fused feature obtained by the fully connected layer belongs to class k; N represents the total number of fused features, K is the number of categories; F′ wn represents the expected feature vector of the synchronous waveform data in the nth feature channel; F wn Represents the feature vector extracted from the nth feature channel of the synchronous waveform data.

[0079] The above feature loss function comprehensively considers classification and regression losses, and the total loss function can help the model obtain more accurate feature representation in the fault detection process, thereby improving the overall performance.

[0080] We calculate the loss value to perform gradient derivation on the network parameters (such as convolution kernel parameters, weights, etc.). We use optimization algorithms (such as gradient descent or Adam) to adjust the network parameters according to the gradient to reduce the loss value. By repeatedly performing the above steps and continuously optimizing the network parameters through multiple iterations, the loss function value eventually converges to the minimum value and the network reaches the optimal state.

[0081] The training method of the above-mentioned adaptive window convolutional network first uses a first sliding window with a larger sliding window size to calculate the fluctuation characteristics of each sample synchronous waveform data, and then screens out the suspected fault area, thereby locating the fault area and effectively improving the recognition accuracy of the fault area. Then, in the suspected fault area, a second sliding window with a smaller size is used to extract local features, and the local features of different samples are fused point by point by convolution to achieve efficient fusion of cross-sample features. Finally, the extracted local features are converted into feature vectors through a fully connected layer, and then the loss value is calculated by a feature loss function. Through iterative optimization of the loss function, the model is continuously adjusted until convergence, thereby finally outputting high-precision fault features. The present invention solves the problems of low efficiency in fault feature extraction, insufficient recognition accuracy, and insufficient cross-channel information fusion in traditional methods.

[0082] In one embodiment, Figure 3 As shown, a method for extracting features of synchronous waveform data is proposed, and the device includes:

[0083] S301, obtaining a plurality of actually measured synchronous waveform data within a preset time period.

[0084] That is, a plurality of synchronous waveform data are collected from a preset time period, and these data include voltage waveform data, current waveform data and other measured values ​​reflecting the operating status of the power system.

[0085] S302, inputting a plurality of actually measured synchronous waveform data into an adaptive window convolutional network to obtain feature vectors of the plurality of actually measured synchronous waveform data.

[0086] The adaptive window convolutional network is trained by the above method. The input data is processed by the adaptive window convolutional network, and feature vectors representing the key characteristics of the synchronous waveform data are output. These feature vectors can be used for subsequent analysis, such as fault detection or classification tasks.

[0087] Figure 4 FIG. 2 shows an internal structure diagram of a feature extraction device for synchronous waveform data in one embodiment. Figure 4 As shown, the feature extraction device for synchronous waveform data includes a processor, a memory and a network interface connected via a system bus. The memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium of the feature extraction device for synchronous waveform data stores an operating system and may also store a computer program. When the computer program is executed by the processor, the processor may implement the feature extraction method for synchronous waveform data. The internal memory may also store a computer program. When the computer program is executed by the processor, the processor may implement the feature extraction method for synchronous waveform data. Those skilled in the art will understand that Figure 4 The structure shown in the figure is merely a block diagram of a portion of the structure related to the scheme of the present application, and does not constitute a limitation on the feature extraction device for synchronized waveform data to which the scheme of the present application is applied. The specific feature extraction device for synchronized waveform data may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0088] A computer-readable storage medium stores a computer program, which implements the following steps when executed by a processor: obtaining a sample set; wherein the sample set includes multiple samples, each sample includes multiple sample synchronous waveform data within a preset time period; inputting the samples into an adaptive window convolution network in sequence, traversing the fluctuation characteristic value of each sample synchronous waveform data in the calculation window through the first sliding window of the sliding window layer, and screening the suspected fault area of ​​each sample synchronous waveform data based on the fluctuation characteristic value; extracting local features in the suspected fault area of ​​each sample synchronous waveform data through the second sliding window of the depthwise separable convolution layer, and performing point-by-point convolution on the local features of different sample synchronous waveform data to obtain fused features; converting the fused features into feature vectors of synchronous waveform data through a fully connected layer, inputting the feature vectors of the synchronous waveform data into a preset feature loss function to calculate the loss value, and iteratively adjusting the adaptive window convolution network through the loss value until the adaptive window convolution network converges.

[0089] And implement the following steps: obtain multiple measured synchronous waveform data within a preset time period; input the multiple measured synchronous waveform data into an adaptive window convolutional network to obtain feature vectors of the multiple measured synchronous waveform data; wherein the adaptive window convolutional network is trained by the above-mentioned adaptive window convolutional network training method.

[0090] A feature extraction device for synchronous waveform data comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the following steps when executing the computer program: obtaining a sample set; wherein the sample set comprises a plurality of samples, each of which comprises a plurality of sample synchronous waveform data within a preset time period; sequentially inputting the samples into an adaptive window convolution network, traversing and calculating the fluctuation characteristic value of each sample synchronous waveform data within the window through a first sliding window of a sliding window layer, and screening the suspected fault area of ​​each sample synchronous waveform data based on the fluctuation characteristic value; extracting local features in the suspected fault area of ​​each sample synchronous waveform data respectively through a second sliding window of a depth-separable convolution layer, and performing point-by-point convolution on the local features of different sample synchronous waveform data to obtain fused features; converting the fused features into feature vectors of synchronous waveform data through a fully connected layer, inputting the feature vectors of the synchronous waveform data into a preset feature loss function to calculate a loss value, and iteratively adjusting the adaptive window convolution network through the loss value until the adaptive window convolution network converges.

[0091] And implement the following steps: obtain multiple measured synchronous waveform data within a preset time period; input the multiple measured synchronous waveform data into an adaptive window convolutional network to obtain feature vectors of the multiple measured synchronous waveform data; wherein the adaptive window convolutional network is trained by the above-mentioned adaptive window convolutional network training method.

[0092] It should be noted that the above-mentioned method, device, equipment and computer-readable storage medium for extracting features from synchronized waveform data belong to a general inventive concept, and the contents of the embodiments of the method, device, equipment and computer-readable storage medium for extracting features from synchronized waveform data are applicable to each other.

[0093] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0094] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0095] The above embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A training method for an adaptive window convolutional network, characterized in that: The adaptive window convolution network includes a sliding window layer, a depthwise separable convolution and a fully connected layer connected in sequence, the sliding window layer is provided with a first sliding window of a first size, the depthwise separable convolution layer is provided with a second sliding window of a second size, the first size is larger than the second size, and the method includes: Acquire a sample set; wherein the sample set includes a plurality of samples, each sample including a plurality of sample synchronous waveform data within a preset time period; Input the samples into the adaptive window convolution network in sequence, traverse and calculate the fluctuation characteristic value of each sample synchronization waveform data in the window through the first sliding window of the sliding window layer, and screen the suspected fault area of ​​each sample synchronization waveform data based on the fluctuation characteristic value; Extracting local features in the suspected fault area of ​​each sample synchronous waveform data through the second sliding window of the depthwise separable convolution layer, and performing point-by-point convolution on the local features of different sample synchronous waveform data to obtain fused features; The fused features are converted into feature vectors of synchronized waveform data through the fully connected layer, the feature vectors of synchronized waveform data are input into a preset feature loss function to calculate a loss value, and the adaptive window convolutional network is iteratively adjusted according to the loss value until the adaptive window convolutional network converges.

2. The method according to claim 1, characterized in that The step of traversing and calculating the fluctuation characteristic value of each sample synchronous waveform data in the window through the first sliding window of the sliding window layer, and screening the suspected fault area of ​​each sample synchronous waveform data based on the fluctuation characteristic value, comprises: Calculate the following when each first sliding window is sliding into the window: In the above formula, μ w (t) represents the mean value at time t; σ w (t) represents the standard deviation at time t; κ w (t) represents the kurtosis at time t; W1 represents the first dimension; X i (k) represents the kth data point of the i-th sample synchronous waveform data in the first sliding window; A window whose mean is greater than a preset first threshold, and / or whose standard deviation is greater than a preset second threshold, and / or whose kurtosis is greater than a preset third threshold is regarded as a suspected fault area.

3. The method according to claim 1, characterized in that The first sliding window of the sliding window layer is used to traverse and calculate the fluctuation characteristic value of each sample synchronous waveform data in the window, and the suspected fault area of ​​each sample synchronous waveform data is screened based on the fluctuation characteristic value, which is expressed as: Calculate the following when each first sliding window is sliding into the window: In the above formula, γ w (t) represents the skewness at time t; R w (t) represents the extreme value difference at time t; μ XWMU represents the mean; W1 represents the first dimension; μ w (t) represents the mean value at time t; σ w (t) represents the standard deviation at time t; X i (k) represents the kth data point of the i-th sample synchronous waveform data in the first sliding window; A window whose skewness is greater than a preset fourth threshold value and / or whose extreme value difference is greater than a preset fifth threshold value is taken as a suspected fault area.

4. The method according to claim 1, characterized in that: The second sliding window of the depthwise separable convolutional layer extracts local features in the suspected fault area of ​​each sample synchronization waveform data, which is expressed as: In the above formula, Y in represents the local features extracted from the nth suspected fault area in the i-th sample synchronous waveform data; X in (t) represents the tth data point in the nth suspected fault area in the i-th sample synchronous waveform data; W2(K) represents the convolution kernel of the second sliding window, with a size of K; X in (t+k) represents the data value at the t+kth time point in the suspected fault area; W2(k) represents the kth element of the convolution kernel.

5. The method according to claim 1, characterized in that The local features of the synchronous waveform data of different samples are convolved point by point to obtain fusion features, which are expressed as: In the above formula, F n represents the nth fusion feature, C represents the total number of sample synchronization waveform data in each sample, and Y in represents the local features extracted from the nth suspected fault area in the i-th sample synchronous waveform data, W 1×1 Represents a point-wise convolution kernel of size 1×1.

6. The method according to claim 1, characterized in that The feature loss function is expressed as: L total =λ CE ·L CE +λ MSE ·L MSE In the above formula, L total represents the loss value; λ CE , MSE is a hyperparameter; L CE is the classification loss; L MSE is the regression loss; is the one-hot encoding form of the true label of the nth sample, where k is the index of the class; is the predicted probability that the nth fused feature obtained through the fully connected layer belongs to class k; N represents the total number of fused features, K is the number of categories; F w ' n represents the expected feature vector of the synchronous waveform data in the nth feature channel; F wn Represents the feature vector extracted from the nth feature channel of the synchronous waveform data.

7. A method for extracting features from synchronous waveform data, characterized in that: The method comprises: Acquire multiple measured synchronous waveform data within a preset time period; The multiple measured synchronous waveform data are input into an adaptive window convolutional network to obtain feature vectors of the multiple measured synchronous waveform data; wherein the adaptive window convolutional network is trained by the method as described in claim 1.

8. A computer-readable storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 7.

9. A feature extraction device for synchronous waveform data, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 7.

Images