Processing monitoring data filling method based on channel attention and temporal convolutional network

Abstract

The present invention provides a method for filling in processing monitoring data based on channel attention and a temporal convolutional network. The method comprises: preprocessing the processing monitoring data to obtain feature samples of the same length; extracting correlation features from the feature samples using a channel attention mechanism; mining the temporal features of the correlation features using a temporal convolutional network to obtain temporal variation features; and converting the temporal variation features into a data space to obtain filled processing monitoring data. This method utilizes historical information from processing monitoring data to predict missing information and mines the inherent connections between the data in sensor and time dimensions. This overcomes the problem of existing techniques that fill in only a few dimensions and lack analysis from the perspective of combining multidimensional data correlation and temporal variation.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method for filling in processing monitoring data based on channel attention and time domain convolutional network. Background Art

[0002] In the field of tool remaining life prediction research, tool processing monitoring data can identify tool processing status and effectively facilitate tool changes during the machining process, preventing premature tool changes that lead to reduced tool utilization or late tool changes that lead to reduced part quality. With the development of industrial big data technology, the manufacturing industry is undergoing a digital transformation and upgrade, generating a large amount of processing monitoring data daily. However, various complex factors on the production site, such as vibration during the machining process causing sensor loosening, noise, and network environment causing data loss, can significantly affect the acquisition of processing monitoring data, leading to data loss and posing challenges to subsequent tool processing status identification and remaining tool life prediction.

[0003] Current shortcomings:

[0004] 1. Tool machining monitoring data suffers from missing data due to improper sensor installation, resulting in missing data in all dimensions at the current point in time. Existing data filling methods mostly focus on filling in missing dimensions in the data, making them unsuitable for tool machining monitoring data.

[0005] 2. Existing time series prediction methods for filling missing data lack dimensional correlation analysis of multidimensional time series data

[0006] After searching, the Chinese patent application number CN202011500119.7 discloses a method and system for online filling of MES system data based on LSTM network. This patent only extracts features of the time dimension of historical information data through the LSTM network, and does not fill in the multiple dimensions of the data itself.

[0007] A Chinese patent application numbered CN202010158581.7 discloses a method and related device for filling missing distribution network voltage data. However, the method lacks the characteristics of historical voltage changes over time, and the patent requires similarity calculation of some attributes to obtain historical voltage data, which makes it unsuitable for filling completely missing data. Summary of the Invention

[0008] In view of the defects in the prior art, the purpose of the present invention is to provide a method for processing monitoring data filling based on channel attention and time domain convolutional network.

[0009] According to one aspect of the present invention, a method for filling in processing monitoring data based on channel attention and time domain convolutional network is provided, comprising:

[0010] Preprocess the processing monitoring data to obtain feature samples of the same length;

[0011] Extracting channel correlation features of the feature samples using a channel attention mechanism;

[0012] Using a time domain convolutional network to mine the channel correlation features, a time series feature is obtained.

[0013] The time series variation characteristics are converted into data space to obtain filled processing monitoring data.

[0014] Preferably, the preprocessing of the processing monitoring data to obtain characteristic samples of the same length includes:

[0015] The processing monitoring data at each moment includes multiple sensor dimensions, and the mean-variance normalization is performed on each of the sensor dimensions;

[0016] For the missing moment to be predicted, all the standardized processing monitoring data before the missing moment are spliced ​​in chronological order, and the data before the missing moment are repeatedly spliced ​​until the length of the processing monitoring data reaches the set value, that is, feature samples of the same length are obtained.

[0017] Preferably, the extracting the channel correlation features of the feature samples by using a channel attention mechanism includes:

[0018] The input feature samples are subjected to global average pooling, the size dimension of the feature samples is compressed to 1, and the channel dimension remains unchanged;

[0019] Adaptively design a one-dimensional convolutional layer, defined as:

[0020] Among them, kernelSize represents the size of the one-dimensional convolution kernel, c represents the number of channels of the input feature sample, and b and γ are both custom parameters;

[0021] Using the one-dimensional convolution kernel to perform a convolution operation on the feature samples after the global average pooling, to obtain the weight of each channel dimension;

[0022] Normalizing the weight of each channel dimension;

[0023] The normalized weights and the input feature samples are multiplied channel by channel to obtain channel correlation features, that is, weighted hidden feature variables.

[0024] Preferably, the using of a time domain convolutional network to perform time series feature mining on the channel correlation features to obtain time series variation features includes:

[0025] The channel correlation features are input into the time domain convolutional network, and the time series variation features of the processed monitoring data are output; wherein the time series convolutional network includes multiple residual modules connected in series; each of the residual modules includes an identity mapping and multiple hidden layers.

[0026] Preferably, in the hidden layer, the temporal variation characteristics of the input data are mined sequentially through dilated causal convolution calculation, network weight normalization, LeakyReLU activation function calculation, and Dropout mask layer;

[0027] The calculation process of the dilated causal convolution, network weight normalization and LeakyReLU activation function is as follows:

[0028] u l+1 =σ(g((b l +Conv1D(w l ,h,s l ,d l ,p l )))

[0029]

[0030] Among them, l is the number of residual layers, w l is the convolution kernel weight factor of the lth layer, s l is the size of the convolution kernel in layer l, d l is the dilation factor of the lth convolution layer, p l is the filling factor of the lth layer of convolution, b l is the bias factor of the convolution kernel of the lth layer; g(·) is the normalization operation on the weight w of the convolution layer, where a is the learnable coefficient after normalization, v is the normalized weight vector, and σ(·) is the LeakyReLU activation function;

[0031] The calculation process of the Dropout mask layer is:

[0032]

[0033] M=(m1,m2,...,m c )

[0034] m j ~Bernoulli(1-r)

[0035] is the result of the calculation through the mask layer, M is the mask vector of the lth layer, the residual layer has a total of c nodes with weight parameters, and the element m j Represents the activation state of the node, and the probability of being 1 is 1-r, that is, the output of each node is randomly set to zero with probability r.

[0036] Preferably, the converting of the time series variation characteristics into the data space to obtain the filled processing monitoring data includes:

[0037] Set a regression module at the end of the time domain convolutional network;

[0038] The time series variation features are converted into data space through the regression module, and the filled processing monitoring data are finally output.

[0039] Preferably, the model is trained by processing the deviation of the observation value of the monitoring data, and the loss function is:

[0040]

[0041] Where y i is the i-th real processing monitoring data, is the i-th estimated processing monitoring data, and n is the number of data.

[0042] According to a second aspect of the present invention, a processing monitoring data filling system based on channel attention and temporal convolutional network is provided, comprising:

[0043] The data processing module pre-processes the processing monitoring data to obtain feature samples of the same length;

[0044] An attention mechanism module, which uses a channel attention mechanism to extract the correlation features of the feature samples;

[0045] A time domain convolution module uses a time domain convolution network to perform time series feature mining on the correlation features to obtain time series change features;

[0046] The filling module converts the time series variation characteristics into data space to obtain filled processing monitoring data.

[0047] According to a third aspect of the present invention, a terminal is provided, comprising a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the program, it can be used to execute the processing monitoring data filling method based on channel attention and time domain convolutional network, or to run the processing monitoring data filling system based on channel attention and time domain convolutional network.

[0048] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the program is executed by a processor, it can be used to execute the processing monitoring data filling method based on channel attention and time domain convolutional network, or to run the processing monitoring data filling system based on channel attention and time domain convolutional network.

[0049] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0050] The processing monitoring data filling method and system based on channel attention and time domain convolutional network in the embodiment of the present invention can use the historical information of processing monitoring data to predict missing information and explore the intrinsic connection of data in sensor dimension and time dimension, overcoming the problems of existing technologies that the filling scope is limited to some dimensions and lacks analysis from the perspective of combining multidimensional data correlation and time changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0052] Figure 1 Flowchart of a method for filling tool processing monitoring data based on channel attention and time-domain convolutional network in one embodiment of the present invention;

[0053] Figure 2 Schematic diagram of the construction process of input feature samples in a preferred embodiment of the present invention;

[0054] Figure 3 1 is a computational logic diagram of a channel attention network (ECA) in a preferred embodiment of the present invention;

[0055] Figure 4 1 is a computational logic diagram of a time-domain convolutional network (TCN) in a preferred embodiment of the present invention;

[0056] Figure 5 The MAE image and RMSE image are compared between the method of the present invention and other methods in a specific application example. DETAILED DESCRIPTION

[0057] The present invention will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those skilled in the art, several variations and improvements can be made without departing from the scope of the present invention. These all fall within the scope of protection of the present invention.

[0058] During the manufacturing process, the remaining life prediction of the tool plays an important role in improving the tool utilization efficiency and part processing quality. The tool processing monitoring data is the data source in the tool remaining life prediction method, and its integrity and effectiveness are crucial.

[0059] In actual processing and production sites, sensor failure or loosening of installed sensors due to vibrations generated by the tool during processing can disrupt the collection of tool processing monitoring data. Many downstream methods and analyses require complete data support in practice, so data filling is of great significance for downstream data analysis.

[0060] The present invention adopts channel attention network to mine the correlation between multiple dimensions in multidimensional time series data to obtain channel correlation feature variables, then uses time domain convolutional network to mine the time variation characteristics of channel correlation feature variables, and finally uses fully connected network to transform the mined hidden feature variables into data space to obtain the final estimated value.

[0061] Based on the above invention concept, the present invention provides an embodiment, see Figure 1 , a processing monitoring data filling method based on channel attention and time domain convolutional network, specifically for tool processing monitoring data, the specific process is as follows:

[0062] S100, pre-processing the tool processing monitoring data to obtain feature samples with the same length;

[0063] S200, uses the channel attention mechanism to extract the channel correlation features of the feature samples preprocessed by S100;

[0064] S300, using a time domain convolutional network to perform time series feature mining on the channel correlation features obtained in S200 to obtain time series change features;

[0065] S400, converting the time series variation characteristics obtained in S300 into a data space to obtain filled tool processing monitoring data.

[0066] This embodiment uses historical information of tool processing monitoring data to predict missing information and mines the intrinsic connection between data in sensor dimension and time dimension, overcoming the problems of existing technologies that the filling scope is limited to some dimensions and lacks analysis from the perspective of combining multidimensional data correlation and time changes.

[0067] In a preferred embodiment of the present invention, S100 is implemented to pre-process the tool processing monitoring data to obtain feature samples of the same length. The specific process includes:

[0068] S101: The tool processing monitoring data at each moment includes multiple sensor dimensions, and the mean and variance of each sensor dimension are normalized. Specifically, the tool processing monitoring data is z-score normalized and the values ​​of the data in different dimensions are processed. The formula is as follows:

[0069]

[0070] Among them, z i Represents the value after standardization, x i Represents the original value, N represents the total number of data, Indicates the average value of all values ​​in a dimension.

[0071] S102: For the missing moment to be predicted, all tool processing monitoring data before the missing moment after being normalized in S101 are spliced ​​in chronological order, and the data before the missing moment are repeatedly spliced ​​until the length of the tool processing monitoring data reaches the set value. Specifically:

[0072] The input data sample size is set to m dimensions, representing sensor feature values ​​of different dimensions. Assuming the data sample time to be predicted is T+1, the input feature sample must include all data samples at time t=0, 1, 2, ..., T. Due to the different processing times of different tools, the number of processing monitoring data included varies. The maximum number of processing monitoring data included across all tools is set to the time step size of the input TCN feature sample, as shown in the following formula:

[0073] s=max(t lk ),l=0,1,2,...,L;k=0,1,2,...,K l

[0074] Among them, l represents the tool number, k represents the time value of tool processing, t lk It represents the ranking of the processing monitoring data obtained by tool l at time k among all the processing monitoring data of tool l.

[0075] According to the above formula, the number of feature samples composed of the tool data samples at time T is less than or equal to the time step s. Therefore, the data samples at the last moment are used to fill the time step of the feature samples. Figure 2 Indicates the construction process of feature samples. The completed feature samples are as follows:

[0076] X T =[x0,x1,x2,...,x T ,...,x T ] s×m

[0077] Wherein, x represents a tool processing monitoring data sample, which is a one-dimensional vector with a length of m.

[0078] In this embodiment, the input feature samples constructed by the above method have two functions: on the one hand, the length of the samples is unified to adapt to the training requirements of the channel attention network and the time domain convolutional network; on the other hand, the moment before the missing moment is closest to the missing data, and repeatedly splicing the processed monitoring data before the missing moment can increase the attention of the data at the previous moment.

[0079] In a preferred embodiment of the present invention, S200 is implemented, and a channel attention network is introduced to calculate the correlation of each dimension of the multidimensional processing monitoring data to obtain the weight of each dimension, and then weighted calculation is performed on the data on each dimension to obtain the channel correlation characteristics of the multidimensional data as a hidden variable output.

[0080] See also Figure 3 , the specific process is:

[0081] S201, average pooling: The tool processing monitoring data at the current prediction time t, the input feature sample is On the time scale, it is the set time step size s; on the sensor dimension, it is m, which represents the various attribute data observed by the sensor; the input feature sample After global average pooling, the time scale of feature samples is compressed to 1, while the sensor dimension remains unchanged.

[0082]

[0083] S202, one-dimensional convolution kernel: Adaptively design a one-dimensional convolution layer, perform convolution operations on the pooled features, calculate the correlation features of each sensor dimension in the data, and output the weight vector on the sensor dimension. The convolution kernel calculation formula is as follows.

[0084]

[0085] Where kernelSize represents the calculated size of the one-dimensional convolution kernel, c represents the number of channels of the input feature sample, b and γ are parameters, which are 1 and 2 respectively in this embodiment. The calculation formula for one-dimensional convolution is as follows:

[0086]

[0087] in, It is the weight data on the kth sensor dimension after the convolution operation, is the i-th pooling eigenvalue, and V[ki] is the weight parameter at the ki position on the convolution kernel.

[0088] S203, calculate channel correlation features: multiply the normalized weights and the original input feature samples channel by channel to obtain the correlation features of the multidimensional data, that is, the weighted hidden feature variables The formula for channel-by-channel multiplication is as follows:

[0089]

[0090] In this embodiment, channel attention is adopted to realize the extraction of channel correlation features of multiple dimensions in multi-dimensional sensor monitoring data, and solve the correlation analysis problem of lack of dimensional data in the process of filling multi-dimensional monitoring data.

[0091] In a preferred embodiment of the invention, S300 is implemented to perform time-scale feature mining on the hidden feature variables output by the channel attention network based on the time domain convolutional network to capture the time variation trend of the time series data. The calculation logic diagram is as follows: Figure 4 Specifically:

[0092] Based on the hidden feature variables calculated by the channel attention network in S200 above, The data of s time steps are input into the time domain convolutional network to mine the changing trend of the data at the previous moment in the subsequent moment.

[0093] from Figure 4 As can be seen in the figure, the time domain convolutional network consists of multiple residual layers. Each residual layer consists of a hidden layer F(x) and an identity mapping x. That is, the residual layer adds the identity mapping x to the hidden layer F(x) based on the input features, which allows the network to not degrade as the depth increases. The formula for a single residual layer is as follows:

[0094] G(x)=σ(F(x)+x)

[0095] Where x is the input, G(x) is the output of the residual layer, F(x) is the calculation result of the hidden layer, and x is the identity mapping.

[0096] In each residual layer, multiple hidden layers are included, and multiple hidden layers are connected in series. The formula for calculating the output of the next hidden layer based on the previous hidden layer variable of the time domain convolutional network is as follows:

[0097] p z =C(p z-1 )

[0098] Where p z-1 is the output of the z-1th hidden layer, p z is the output of the z-th hidden layer, and C(·) is the network structure of the z-th hidden layer.

[0099] Furthermore, each hidden layer uses an expanded causal convolution layer, a weight normalization layer, a LeakyReLU activation function, and a Dropout mask layer to mine the time scale characteristics of the data. The formula is as follows:

[0100]

[0101]

[0102] M=(m1,m2,...,m c )

[0103] m j ~Bernoulli(1-r)

[0104] Where l is the number of hidden layers, u l+1 is the output of the lth hidden layer and the input of the l+1th hidden layer, w l is the convolution kernel weight factor of the lth layer, s l is the size of the convolution kernel in layer l, d l is the dilation factor of the lth convolution layer, p l is the filling factor of the lth layer of convolution, b l is the bias factor of the convolution kernel of the lth layer. g(·) is the normalization of the weight w of the convolution layer. The weight normalization method is to decouple the weight vector w in the network into the parameter vector v and the parameter scalar in its Euclidean norm and its direction. Then they are optimized separately. σ(·) is the Leaky ReLU activation function. M is the mask vector of the lth layer, M=(m1,m2,...,m c ) indicates that the residual layer has a total of c weight parameters, and the element m j Represents the activation state of the node, Bernoulli (1-r) represents the Bernoulli distribution, that is, m j The probability of being 1 is 1-r, that is, the output of each node is multiplied by m j The probability r is then randomly set to zero.

[0105] In this embodiment, a time-domain convolutional network is adopted to realize the extraction of temporal features of monitoring data using a causal convolutional network, thereby solving the problem of low filling accuracy of temporal monitoring data.

[0106] In a preferred embodiment of the present invention, S400 is implemented to convert the hidden feature variables obtained in S300, i.e., the time-series variation features, into the data space to obtain padded tool processing monitoring data. Specifically, a regression module is provided at the end of the time-domain convolutional network to convert the hidden feature variables into the final output. This embodiment uses a fully connected layer, and the input is an estimate of the next moment.

[0107] Specifically, the time domain convolutional network outputs the hidden feature variable u, which contains the correlation characteristics of the channel attention network and the time domain convolutional network in multiple sensor dimensions, as well as the time change trend characteristics. Finally, the data is converted to the data space through the fully connected layer network to obtain the estimated value of the filling at time t The formula is as follows:

[0108]

[0109] in, represents the filled tool processing monitoring data, u represents the hidden variable output by the time domain convolutional network. q(·) is the dimension change of the hidden variable to adapt to the number of nodes in the fully connected network, w D and b D are the weights and bias values ​​of the fully connected layer, and f(·) represents the activation function.

[0110] In a preferred embodiment of the present invention, the model is trained by processing the deviation of the monitoring data observation value, and the loss function is as follows:

[0111]

[0112] Where y i is the i-th real processing monitoring data, is the i-th estimated processing monitoring data, and n is the number of data.

[0113] Based on the same inventive concept, the present invention also provides a processing monitoring data filling system based on channel attention and time domain convolutional network, specifically tool processing monitoring data, including:

[0114] The data processing module pre-processes the tool processing monitoring data to obtain feature samples of the same length;

[0115] An attention mechanism module, which uses a channel attention mechanism to extract the correlation features of the feature samples;

[0116] A time domain convolution module uses a time domain convolution network to perform time series feature mining on the correlation features to obtain time series change features;

[0117] The filling module converts the time series variation characteristics into the data space to obtain the filled tool processing monitoring data.

[0118] It should be noted that the method in the embodiment of the present invention is not only applicable to tool processing monitoring data, but can also be used to fill in the gaps in time-series multi-dimensional monitoring data in other fields.

[0119] Based on the same inventive concept, in other embodiments of the present invention, a terminal is provided, comprising a memory, a processor, and a computer program stored on the memory and runnable on the processor. When the processor executes the program, it can be used to execute the above-mentioned processing monitoring data filling method based on channel attention and time domain convolutional network, or, run the above-mentioned processing monitoring data filling system based on channel attention and time domain convolutional network.

[0120] Optionally, the memory is used to store programs; the memory may include volatile memory (English: volatile memory), such as random-access memory (English: random-access memory, abbreviated: RAM), such as static random-access memory (English: static random-access memory, abbreviated: SRAM), double data rate synchronous dynamic random access memory (English: Double Data Rate Synchronous Dynamic Random Access Memory, abbreviated: DDRSDRAM), etc.; the memory may also include non-volatile memory (English: non-volatile memory), such as flash memory (English: flash memory). The memory is used to store computer programs (such as applications, functional modules, etc. that implement the above-mentioned methods), computer instructions, etc. The above-mentioned computer programs, computer instructions, etc. can be partitioned and stored in one or more memories. And the above-mentioned computer programs, computer instructions, data, etc. can be called by the processor.

[0121] The processor is configured to execute the computer program stored in the memory to implement the various steps of the method involved in the above embodiment. For details, please refer to the relevant description in the above method embodiment.

[0122] The processor and memory can be independent structures or integrated structures. When the processor and memory are independent structures, the memory and processor can be coupled via a bus.

[0123] Based on the same inventive concept, in other embodiments of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When the program is executed by a processor, it can be used to execute the above-mentioned processing monitoring data filling method based on channel attention and time domain convolutional network, or to run the above-mentioned processing monitoring data filling system based on channel attention and time domain convolutional network.

[0124] Computer-readable media include computer storage media and communication media, wherein communication media include any medium that facilitates the transfer of computer programs from one location to another. The storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer. An exemplary storage medium is coupled to a processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be an integral part of the processor. The processor and storage medium can be located in an ASIC. Alternatively, the ASIC can be located in a user device. Of course, the processor and storage medium can also exist as discrete components in a communication device.

[0125] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0126] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0127] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0128] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0129] In other embodiments of the present invention, the present method and other methods are used to populate tool processing monitoring data. Specifically, the data used is acoustic emission sensor data collected during the machining of wheel slot milling cutters at a steam turbine plant. Four-dimensional monitoring data is used for verification: energy, average frequency, root mean square value, and average signal level.

[0130] The mean absolute error (MAE) and root mean square error (RMSE) images are used to compare the method of the present invention with other methods. Figure 5 .

[0131] Let y i For tool processing monitoring data, To fill the data, N is the number of missing values, m is the number of dimensions of the processed monitoring data, and the definitions of MAE and RMSE are:

[0132]

[0133]

[0134] Method 1 is a forward substitution method, in which the missing values ​​at time t are replaced by the missing values ​​at time t-1. Method 2 is an exponential smoothing method, which regards the predicted value as the weighted sum of the previous observations and gives different weights to different data, with new data given larger weights and old data given smaller weights. Method 3 uses a time-domain convolutional network to mine the temporal characteristics of tool processing monitoring data and predict and fill in the data at the missing positions. It does not consider the correlation of multi-dimensional sensor data in the processing monitoring data. Method 4 ECA-TCN is the algorithm proposed in this paper. In the experiment, the TCN-based processing monitoring data filling method and the ECA-TCN-based processing monitoring data filling method were both tested 10 times, and the average prediction error of the experiment was compared with other methods. According to the experimental results obtained, compared with the above methods, it is proved that the tool processing monitoring data filling method proposed in this invention achieves better results.

[0135] The above describes specific embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various modifications or variations within the scope of the claims without affecting the essence of the present invention. The above preferred features may be used in any combination as long as they do not conflict with each other.

Claims

1. A method for processing monitoring data filling based on channel attention and time domain convolutional network, characterized in that: include: Preprocess the tool processing monitoring data to obtain feature samples of the same length; Extracting channel correlation features of the feature samples using a channel attention mechanism; Using a time domain convolutional network to perform time series feature mining on the channel correlation features to obtain time series change features; Converting the time series variation characteristics into data space to obtain filled tool processing monitoring data; The preprocessing of the tool processing monitoring data to obtain feature samples of the same length includes: The processing monitoring data at each moment includes multiple sensor dimensions, and mean-variance normalization is performed on each of the sensor dimensions; For the missing moment to be predicted, all the standardized processing monitoring data before the missing moment are spliced ​​in chronological order, and the data before the missing moment are repeatedly spliced ​​until the length of the processing monitoring data reaches the set value, that is, feature samples of the same length are obtained; The method of using a time domain convolutional network to perform time series feature mining on the channel correlation feature to obtain time series change features includes: Inputting the channel correlation features into the time domain convolutional network and outputting the time series variation features of the processed monitoring data; wherein the time domain convolutional network includes a plurality of residual modules connected in series; each of the residual modules includes an identity map and a plurality of hidden layers; In the hidden layer, the temporal variation characteristics of the input data are mined through dilated causal convolution calculation, network weight normalization, LeakyReLU activation function calculation, and Dropout mask layer in sequence; The calculation process of the dilated causal convolution, network weight normalization and LeakyReLU activation function is as follows: u l+1 nσ(g((b l +Conv1D(w l ,h,s l ,d l ,p l ))) Among them, l is the number of residual layers, w l is the convolution kernel weight factor of the lth layer, s l is the size of the convolution kernel in layer l, d l is the dilation factor of the lth convolution layer, p l is the filling factor of the lth layer of convolution, b l is the bias factor of the convolution kernel of the lth layer; g(·) is the normalization operation on the weight w of the convolution layer, where a is the learnable coefficient after normalization, v is the normalized weight vector, and σ(·) is the LeakyReLU activation function; The calculation process of the Dropout mask layer is: M=(m1,m2,...,m c ) m j ~Bernoulli(1-r) is the result of the calculation through the mask layer, M is the mask vector of the lth layer, the residual layer has a total of c nodes with weight parameters, and the element m j Represents the activation state of the node, and the probability of being 1 is 1-r, that is, the output of each node is randomly set to zero with probability r.

2. The method for filling in processing monitoring data based on channel attention and time domain convolutional network according to claim 1 is characterized in that: The extracting the channel correlation features of the feature samples by using the channel attention mechanism includes: The input feature samples are subjected to global average pooling, the size dimension of the feature samples is compressed to 1, and the channel dimension remains unchanged; Adaptively design a one-dimensional convolutional layer, defined as: Among them, kernelSize represents the size of the one-dimensional convolution kernel, c represents the number of channels of the input feature sample, and b and γ are both custom parameters; Using the one-dimensional convolution kernel to perform a convolution operation on the feature samples after the global average pooling, to obtain the weight of each channel dimension; Normalizing the weight of each channel dimension; The normalized weights and the input feature samples are multiplied channel by channel to obtain channel correlation features, that is, weighted hidden feature variables.

3. The method for filling in processing monitoring data based on channel attention and time domain convolutional network according to claim 1 is characterized in that: The converting of the time series variation characteristics into a data space to obtain filled tool processing monitoring data includes: A regression module is set at the end of the time domain convolutional network; The time series variation features are converted into data space through the regression module, and the filled processing monitoring data are finally output.

4. The method for filling in processing monitoring data based on channel attention and time domain convolutional network according to claim 1 is characterized in that: The model is trained by processing the deviation of the observation value of the monitoring data, and the loss function is: Where y i is the i-th real processing monitoring data, is the i-th estimated processing monitoring data, and n is the number of data.

5. A processing monitoring data filling system based on channel attention and time domain convolutional network, used to implement the method according to any one of claims 1 to 4, characterized in that: include: The data processing module pre-processes the tool processing monitoring data to obtain feature samples of the same length; An attention mechanism module, which uses a channel attention mechanism to extract the correlation features of the feature samples; A time domain convolution module uses a time domain convolution network to perform time series feature mining on the correlation features to obtain time series change features; The filling module converts the time series variation characteristics into data space to obtain filled tool processing monitoring data.

6. A terminal comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it can be used to perform the method according to any one of claims 1 to 4, or run the system according to claim 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, it can be used to perform the method described in any one of claims 1 to 4, or to run the system described in claim 5.

Images