Industrial time series data anomaly detection method and system based on time-frequency domain feature enhancement

Abstract

The invention relates to an industrial time series data anomaly detection method and system based on time-frequency domain feature enhancement. The method comprises the steps of collecting industrial time series data in an industrial production process; performing time domain processing on the industrial time series data by combining a weighted weight adjustment strategy in a time domain and a data interpolation method; converting the industrial time sequence data subjected to the time domain processing into frequency domain data, performing frequency domain enhancement processing on the frequency domain data, and then converting the frequency domain data back to a time domain to obtain time-frequency domain enhanced industrial time sequence data; and performing anomaly detection on the time-frequency domain enhanced industrial time series data by adopting a preset double-branch neural network model. According to the method, for the mass time series data of any single source, time domain and frequency domain data enhancement processing is performed on the industrial time series data, and anomaly detection is performed by adopting the preset double-branch neural network model, so that the detection difficulty caused by data quality in the existing method is effectively overcome.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to an industrial time series data anomaly detection method and system based on time-frequency domain feature enhancement. Background Art

[0002] Anomaly detection refers to the automatic identification of abnormal phenomena embedded in a large amount of normal data. In the industrial manufacturing field, the production process is accompanied by a large amount of production data. With the vigorous development of sensor technology and Internet of Things technology, it has become possible to identify and judge abnormal patterns that do not conform to expected states by mining and analyzing data.

[0003] The task goal of anomaly detection is to determine which data has characteristic patterns that are significantly different from other normal data. Usually, considering that the outliers deviate far from the average of other values, the traditional method uses data statistics to detect deviations to determine whether an anomaly occurs. In the field of industrial manufacturing, with the emergence of the Internet of Things (IoT), a large amount of time series data with multi-dimensional, network flow, local flow, mixed flow and other different data domains have been generated. The analysis and anomaly detection methods of these time series data in the industrial production process are related to production safety and have important research significance.

[0004] Existing anomaly detection schemes can be divided into two categories. One is to use traditional statistical analysis methods, such as statistical process control (SPC) chart methods such as cumulative sum charts to detect changes in the characteristics of the basic distribution of data to identify anomalies, which requires high accuracy in the collection of original data; the other is to use machine learning (deep learning) methods to train models through data, and time series prediction models are commonly used for comparison and judgment. However, this method is currently not very practical and the detection results are not stable.

[0005] With the development of the Internet, the Internet of Things, communication technology, and storage technology, the time series data that can be collected and stored by various sensors in the industrial manufacturing field has increased rapidly. The abnormal data that may appear in the time series data needs to be discovered in time to help factory managers assess abnormal risks and take corresponding countermeasures. The existing anomaly detection scheme has the following defects when dealing with discrete detection data collected in the industrial field: 1) The data acquisition and transmission speed in the industrial manufacturing process is difficult to meet the signal processing speed requirements. The data acquisition results are discrete and the sampling frequency is low, and the real information cannot be restored, resulting in distorted detection results of traditional anomaly detection schemes. 2) In the application process of machine learning and deep learning anomaly detection schemes, model training is required to determine the model parameters, but the lack of effective training sample data will make it difficult to train the model parameters and cause overfitting, affecting the model performance. The above defects lead to problems such as low detection result accuracy and low analysis efficiency in traditional anomaly detection schemes. Summary of the invention

[0006] In view of this, it is necessary to provide an industrial time series data anomaly detection method based on time-frequency domain feature enhancement to solve the above-mentioned defects of the prior art.

[0007] In order to solve the above problems, in a first aspect, an embodiment of the present invention provides an industrial time series data anomaly detection method based on time-frequency domain feature enhancement, comprising:

[0008] Collect industrial time series data in industrial production processes;

[0009] Combining the weighted weight adjustment strategy and the data interpolation method in the time domain, the industrial time series data is processed in the time domain;

[0010] Converting the industrial time series data after time domain processing into frequency domain data, performing frequency domain enhancement processing on the frequency domain data and converting it back into the time domain to obtain industrial time series data enhanced in the time and frequency domains;

[0011] A preset dual-branch neural network model is used to perform anomaly detection on industrial time series data enhanced in the time and frequency domains.

[0012] Preferably, when collecting industrial time series data in an industrial production process, the method further includes:

[0013] Based on Shannon sampling theorem and prior knowledge, determine the optimal sampling frequency of industrial time series data; wherein the industrial time series data includes voltage, current, temperature, humidity and speed;

[0014] Determine the data capacity that needs to be expanded for industrial time series data based on the optimal sampling frequency.

[0015] Preferably, the weighted weight adjustment strategy and the data interpolation method in the time domain are combined to perform time domain processing on the industrial time series data in the original data set, including:

[0016] According to the characteristics of industrial time series data, a weight value is assigned to each data point of the industrial time series data;

[0017] Considering the weight value of each data point, data interpolation is performed on the industrial time series data.

[0018] Preferably, after performing data interpolation on the industrial time series data, the method further includes:

[0019] The data after data interpolation is optimized; the optimization includes data rotation, data jitter, data scaling and data rearrangement.

[0020] Preferably, the industrial time series data after time domain processing is converted into frequency domain data, and the frequency domain data is subjected to frequency domain enhancement processing and then converted back to the time domain to obtain industrial time series data enhanced in the time and frequency domains, including:

[0021] The industrial time series data after time domain processing is converted into frequency domain data through fast Fourier transform;

[0022] Performing frequency domain enhancement processing on the frequency domain data; wherein the frequency domain enhancement processing includes frequency domain noise addition, frequency masking, frequency shifting and frequency scaling;

[0023] The data that has been enhanced in the frequency domain is converted back to the time domain through inverse fast Fourier transform to obtain industrial time series data enhanced in the time and frequency domains.

[0024] Preferably, the dual-branch neural network model includes a 1D single-layer convolutional neural network and a long short-term memory network;

[0025] The 1D single-layer convolutional neural network and the long short-term memory network share an input layer and an output layer. A single-layer convolutional network, a feature map series connection, a maximum pooling layer, and a Softmax classification regression layer connected in sequence are arranged between the input layer and the output layer of the 1D single-layer convolutional neural network; an LSTM layer, a hidden feature layer, and a Softmax classification regression layer connected in sequence are arranged between the input layer and the output layer of the long short-term memory network.

[0026] Preferably, the method of using a preset dual-branch neural network model to perform anomaly detection on industrial time series data enhanced in the time and frequency domains includes:

[0027] The dual-branch neural network model includes a 1D single-layer convolutional neural network and a long short-term memory network;

[0028] The 1D single-layer convolutional neural network converts the industrial time series data enhanced in the time and frequency domain into two-dimensional data, slices the two-dimensional data through the single-layer convolutional network, extracts graphic features, outputs multiple feature maps, connects the feature maps in series and inputs them into the maximum pooling layer, sends the output of the maximum pooling layer to the Softmax classification regression layer, and outputs the distribution position of the abnormal data;

[0029] The long short-term memory network processes the industrial time series data enhanced in the time and frequency domains through the LSTM layer to obtain the hidden features of the industrial time series data, and obtains the abnormal data prediction results through softmax classification regression;

[0030] For the data to be detected, the anomaly scores are calculated respectively by 1D single-layer convolutional neural network and long short-term memory network, and the weighted sum is used to obtain the final anomaly detection result.

[0031] In a second aspect, an embodiment of the present invention provides an industrial time series data anomaly detection system based on time-frequency domain feature enhancement, comprising:

[0032] Data acquisition module, used to collect industrial time series data in the industrial production process;

[0033] A time domain processing module, used to perform time domain processing on the industrial time series data in combination with a weighted weight adjustment strategy and a data interpolation method in the time domain;

[0034] A frequency domain enhancement module, used to convert the industrial time series data after time domain processing into frequency domain data, perform frequency domain enhancement processing on the frequency domain data and then convert it back to the time domain to obtain industrial time series data enhanced in the time and frequency domains;

[0035] The anomaly detection module is used to perform anomaly detection on industrial time series data enhanced in the time and frequency domains using a preset dual-branch neural network model.

[0036] In a third aspect, the present invention further provides an electronic device, including a memory and a processor, wherein:

[0037] The memory is used to store programs;

[0038] The processor is coupled to the memory and is used to execute the program stored in the memory to implement the steps in the industrial time series data anomaly detection method based on time-frequency domain feature enhancement as described in the embodiment of the first aspect of the present invention.

[0039] In a fourth aspect, the present invention further provides a computer-readable storage medium for storing computer-readable programs or instructions, which, when executed by a processor, can implement the steps in the industrial time series data anomaly detection method based on time-frequency domain feature enhancement as described in the embodiment of the first aspect of the present invention.

[0040] Compared with the prior art, the industrial time series data anomaly detection method and system based on time-frequency domain feature enhancement provided by the present invention have the following beneficial effects:

[0041] 1) The present invention combines the weighted weight adjustment strategy and data interpolation method in the time domain to perform time domain processing on the original industrial time series data, effectively solving the information distortion problem caused by low sampling frequency and data discreteness in the data acquisition process, and providing a more reliable data basis for subsequent frequency domain enhancement and anomaly detection.

[0042] 2) The present invention converts the data after time domain processing into frequency domain and performs frequency domain enhancement processing, which can highlight the frequency domain characteristics of the data and make it easier for the model to capture abnormal signals. By converting back to the time domain through inverse fast Fourier transform, industrial time series data with enhanced time and frequency domains is obtained, which provides richer feature information for anomaly detection.

[0043] 3) The present invention adopts a preset dual-branch neural network model, which combines the advantages of convolutional neural networks in feature extraction and the expertise of long short-term memory networks in processing time series data, and can simultaneously capture the spatial and temporal characteristics of industrial time series data. This dual-branch structure enables the model to analyze data more comprehensively and improve the accuracy and efficiency of anomaly detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A flow chart of the method for detecting anomalies in industrial time series data based on time-frequency domain feature enhancement provided by the present invention;

[0045] Figure 2 A structural diagram of a dual-branch neural network model provided by the present invention;

[0046] Figure 3 A structural block diagram of the industrial time series data anomaly detection system based on time-frequency domain feature enhancement provided by the present invention;

[0047] Figure 4 This is a structural block diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0048] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0049] 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.

[0050] At present, the existing anomaly detection schemes can be divided into two categories. One is to use traditional statistical analysis methods, such as statistical process control (SPC) chart methods such as cumulative sum charts to detect changes in the characteristics of the basic distribution of data to identify anomalies, which requires high accuracy in the collection of original data; the other is to use machine learning (deep learning) methods to train models through data, and time series prediction models are commonly used for comparison and judgment, but the current method is not practical and the detection results are unstable.

[0051] Existing anomaly detection solutions have the following defects when dealing with discrete detection data collected in the industrial field: 1) The data acquisition and transmission speed in the industrial manufacturing process cannot meet the signal processing speed requirements. The data acquisition results are discrete and the sampling frequency is low, which cannot restore the real information, resulting in distorted detection results of traditional anomaly detection solutions. 2) In the application process of machine learning and deep learning in anomaly detection solutions, model training is required to determine the model parameters, but insufficient effective training sample data will make model parameter training difficult and cause overfitting, affecting model performance. The above defects lead to problems such as low detection result accuracy and low analysis efficiency in traditional anomaly detection solutions.

[0052] In view of this, the present invention provides an industrial time series data anomaly detection method based on time-frequency domain feature enhancement. For massive time series data from any single source, the industrial time series data is enhanced in the time domain and frequency domain, and then a preset dual-branch neural network model is used for anomaly detection, which effectively overcomes the detection difficulties caused by data quality in existing methods. The following will be explained and introduced through multiple embodiments.

[0053] Figure 1 The flowchart of the method for detecting anomalies of industrial time series data based on time-frequency domain feature enhancement provided by the present invention is as follows. Figure 1 As shown, the industrial time series data anomaly detection method based on time-frequency domain feature enhancement includes the following steps S1 to S4, wherein:

[0054] Step S1, collecting industrial time series data in the industrial production process.

[0055] Specifically, in the industrial production process, sensors deployed on the equipment collect time series data of detection indicators such as voltage, current, temperature, humidity, speed, etc. to obtain the original data set. These data are usually recorded in discrete form, and each data point contains a timestamp and the corresponding measurement value. T , which can be defined as

[0056] X T ={x 1 , x 2 , …, x i}

[0057] Where, X T Represents the measured value of the detection index X at time T, x i is the measured value of node i in the industrial time series data,

[0058] When collecting industrial time series data in an industrial production process, the optimal sampling frequency of the industrial time series data is determined based on Shannon's sampling theorem and prior knowledge; wherein the industrial time series data includes voltage, current, temperature, humidity and speed.

[0059] It is understandable that Shannon's sampling theorem states that in order to ensure the complete recovery of the original continuous-time signal, the sampling frequency must be at least twice the highest frequency of the signal. In industrial production, the highest frequency of the signal is usually determined by equipment characteristics and process requirements.

[0060] In this embodiment, the highest frequency fH of the signal can be obtained based on prior knowledge, that is, the highest frequency of the signal can be estimated based on the historical data of the equipment or process and expert experience. According to Shannon sampling theorem, the maximum frequency fH of the signal can be obtained to ensure the effective reconstruction of industrial time series data X T The ideal sampling frequency f, where f≥2fH, is set as the optimal sampling frequency f=2fH for industrial time series data in this embodiment, which is a constant parameter.

[0061] In a preferred embodiment of the present invention, after determining the optimal sampling frequency of the industrial time series data, the data capacity that needs to be expanded for the industrial time series data can be determined based on the optimal sampling frequency.

[0062] Specifically, for the discrete sparse time series data X that has been collected T ={x 1 , x 2 , …, x i}, the original data length is N, and the discrete sampling time series index is T = {1, 2, ..., t}. In order to meet the requirements of the optimal sampling frequency, it is necessary to expand the data capacity. According to the requirements of the ideal sampling frequency f, the newly synthesized data through the data enhancement method is defined as Its data length is expanded to N * =2N*f-f+1, the expanded time series index becomes in:

[0063]

[0064] In this embodiment, after determining the optimal sampling frequency of industrial time series data, the discrete sparse time series data in the original data set is expanded to a data capacity that meets the requirements of the optimal sampling frequency based on the optimal sampling frequency, providing a richer data foundation for subsequent data enhancement.

[0065] Step S2, combining the weighted weight adjustment strategy and data interpolation method in the time domain, to perform time domain processing on the industrial time series data.

[0066] Step S2 may include the following steps S21 to S23, wherein:

[0067] Step S21, assigning a weight value to each data point in the original data set according to the characteristics of the industrial time series data in the original data set.

[0068] In this embodiment, different weights may be assigned according to the time distance or importance of the data points. In this embodiment, the weights of the data points are assigned according to the time distance of the data points. Recent data points are assigned higher weights because they are more able to reflect the current trend or status.

[0069] Step S22, considering the weight value of each data point, and performing data interpolation on the industrial time series data in the original data set.

[0070] In this embodiment, a variety of data interpolation methods such as linear interpolation, polynomial interpolation, and Lagrange interpolation can be used. When performing data interpolation, the weight of the data point needs to be considered. For example, when using linear interpolation or polynomial interpolation, the coefficients or parameters of the interpolation formula can be adjusted according to the weight of the data point to obtain a more accurate interpolation result.

[0071] The embodiment of the present invention combines the weighted weight adjustment strategy and data interpolation method in the time domain to perform time domain processing on the original industrial time series data, effectively solving the information distortion problem caused by low sampling frequency and data discreteness in the data acquisition process, and providing a more reliable data basis for subsequent frequency domain enhancement and anomaly detection.

[0072] Step S23, optimizing the data after weighted interpolation; the optimization processing includes data rotation, data dithering, data scaling and data rearrangement.

[0073] Specifically, in order to further improve the diversity of data, image preprocessing methods in computer vision (such as rotation, jitter, scaling, reordering, etc.) are introduced into the optimization processing of industrial time series data.

[0074] In this embodiment, the intermediate data generated by interpolation For optimization processing, assume that the data before industrial time series data optimization processing is The optimized data is X′ T ;

[0075] Intermediate data generated by data interpolation Four data operations are performed: data rotation, data dithering, data scaling, and data rearrangement, among which:

[0076] Data rotation operation:

[0077]

[0078] Where R is the rotation parameter set.

[0079] Data jitter operation:

[0080]

[0081] Where, ∈ is the set of jitter parameters, ∈ t represents the sum of the superimposed noise at each time t.

[0082] Data scaling operations:

[0083]

[0084] Where W is the amplitude scaling operation parameter, is a sequence, when they are a uniform value α, the data are scaled synchronously; or it can be obtained by cubic spline interpolation method, first through a normal distribution with a standard deviation of σ To select the node u = {u 1 ,u 2 ,…,u t}, and then use cubic spline interpolation CS(u) to find α t :

[0085]

[0086] Data rearrangement operation:

[0087]

[0088] Where i, j, k represent the first index slice of each window, so each slice is selected only once, using slices of uniform window size, w represents the slice window size, Where n is the total number of slices.

[0089] In the intermediate data In the process of performing the four data optimization processing operations of data rotation, data dithering, data scaling and data rearrangement, each data optimization processing operation has two possibilities of being used / not used, with a total of 16 possibilities.

[0090] In this embodiment, the industrial time series data after weighted interpolation is optimized, thereby improving the diversity of the data while retaining the data characteristics.

[0091] Step S3, converting the industrial time series data after time domain processing into frequency domain data, performing frequency domain enhancement processing on the frequency domain data and converting it back to the time domain to obtain industrial time series data enhanced in the time and frequency domains.

[0092] Specifically, first, the industrial time series data after time domain processing is converted into frequency domain data through fast Fourier transform;

[0093] Next, the frequency domain data is subjected to frequency domain enhancement processing; the frequency domain enhancement processing may include frequency domain noise addition, frequency masking, frequency shifting and frequency scaling. Among them, frequency domain noise addition refers to adding random noise in the frequency domain to increase the complexity and diversity of the data. Frequency masking refers to randomly selecting a part of the frequency components in the frequency domain to set them to zero or reduce their amplitude to simulate the attenuation or noise interference of the signal during transmission. Frequency shifting refers to shifting the entire frequency domain feature to the left or right by a certain amount to simulate the delay or advance of the signal in time. Frequency scaling refers to scaling the frequency domain features to simulate the performance of the signal at different speeds.

[0094] Finally, the frequency-domain enhanced data is converted back to the time domain through inverse fast Fourier transform to obtain industrial time series data enhanced in the time and frequency domain. The enhanced data set is recorded as

[0095] The embodiment of the present invention converts the data after time domain processing into the frequency domain and performs frequency domain enhancement processing, which can highlight the frequency domain characteristics of the data and make it easier for the model to capture abnormal signals. By converting back to the time domain through inverse fast Fourier transform, industrial time series data with enhanced time and frequency domains is obtained, which provides richer feature information for anomaly detection.

[0096] Step S4, using a preset dual-branch neural network model to perform anomaly detection on the industrial time series data enhanced in the time and frequency domains.

[0097] Specifically, refer to Figure 2 , the dual-branch neural network model includes a 1D single-layer convolutional neural network and a long short-term memory network;

[0098] The 1D single-layer convolutional neural network and the long short-term memory network share an input layer and an output layer. A single-layer convolutional network, a feature map concatenation (FeatureMap Concat), a maximum pooling layer MaxPooling and a Softmax classification regression layer connected in sequence are arranged between the input layer and the output layer of the 1D single-layer convolutional neural network; an LSTM (Long Short-Term Memory, long short-term memory network) layer, a hidden feature layer (Hidden Layer) and a Softmax classification regression layer connected in sequence are arranged between the input layer and the output layer of the long short-term memory network.

[0099] Among them, the 1D Single-Layer Convolutional Neural Network is a convolutional neural network structure specifically used to process one-dimensional data. Since only one layer of convolution is used, the computational complexity of the model is relatively low, which is suitable for processing large-scale time series data and meeting real-time requirements.

[0100] In this embodiment, the 1D single-layer convolutional neural network converts the industrial time series data enhanced in the time-frequency domain into a time series data with the horizontal axis being time T and the vertical axis being The two-dimensional data is sliced ​​through a single-layer convolutional network, the graphic features are extracted, and multiple feature maps are output. The feature maps are connected in series and input into the maximum pooling layer. The output of the maximum pooling layer is sent to the Softmax classification regression layer to output the distribution position of the abnormal data.

[0101] The long short-term memory network processes the industrial time series data enhanced in the time and frequency domains through the LSTM layer to obtain the hidden features of the industrial time series data, and obtains the abnormal data prediction results through softmax classification regression.

[0102] In the training process of the dual-branch neural network model, the results of the two branches are combined and trained together, which helps the model learn both local and global features at the same time.

[0103] For 1D single-layer convolutional neural network, the root mean square error is used as the loss function:

[0104]

[0105] In the formula, is the loss function of a 1D single-layer convolutional neural network, M is the sample size of the training dataset, and z i It is the prediction result obtained by 1D single-layer convolutional neural network.

[0106] For long short-term memory networks, data similarity error is used as the loss function:

[0107]

[0108] In the formula, is the loss function of the LSTM network. The simi(·) operator calculates {x i}, The normalized covariance between the data, It is the prediction result obtained by the long short-term memory network.

[0109] The loss function of the dual-branch neural network model training is obtained comprehensively for:

[0110]

[0111] In this embodiment, for the data to be detected, the anomaly scores are calculated respectively by a 1D single-layer convolutional neural network and a long short-term memory network, and the weighted sum is performed to obtain the final anomaly detection result.

[0112] Specifically, after the model training is completed, the anomaly score S of the dual-branch neural network model is AD To determine whether the data is abnormal:

[0113] Among them, the data x to be judged i , obtain the sequence data with a surrounding window size of W,

[0114] Prediction result z obtained for 1D single-layer convolutional neural network i , its abnormal score result S A for:

[0115]

[0116] Prediction results for LSTM networks Its abnormal score result S D for:

[0117]

[0118] Finally, the anomaly score S of the two-branch neural network model AD for:

[0119] S AD =λS A +(1-λ)S D

[0120] Where λ is the weight of the 1D single-layer convolutional neural network anomaly detection score, λ∈[0,1]. AD If the score is greater than the preset threshold, it is considered that a data anomaly has occurred and an early warning is issued.

[0121] The present invention adopts a preset dual-branch neural network model, which combines the advantages of convolutional neural networks in feature extraction and the expertise of long short-term memory networks in processing time series data, and can simultaneously capture the spatial and temporal characteristics of industrial time series data. This dual-branch structure enables the model to analyze data more comprehensively and improve the accuracy and efficiency of anomaly detection.

[0122] In a preferred embodiment of the present invention, between step S3 and step S4, the industrial time series data anomaly detection method based on time-frequency domain feature enhancement provided by the present invention further includes:

[0123] The strategy of neighboring point label aggregation is adopted to migrate the label information of the original data set to the data set enhanced in the time and frequency domain by calculating the time domain aggregation feature index.

[0124] Specifically, for the data element x in the original industrial time series dataset i , data element x iThe time domain aggregation feature index during training is expressed as:

[0125]

[0126] Among them, N i is the neighborhood of node i in industrial time series data, x j is the neighborhood value, LeakyReLU(·) is a nonlinear activation function used in the training process, a ij is the aggregate score between data elements i and j;

[0127] Neighborhood normalization is performed through the softmax function to calculate the aggregation score of each neighborhood point;

[0128]

[0129] Among them, d(x i ,x j ) is the data element x i ,x j The distance between

[0130] According to the aggregated scores, the labels in the original data are transferred to the enhanced data points in the time and frequency domains.

[0131] Furthermore, the migrated label information is combined with the time attention mechanism to train a 1D single-layer convolutional neural network model. In this embodiment, taking into account the requirements of industrial application fields for real-time response, a 1D single-layer convolutional neural network model with a time attention mechanism is used to achieve rapid industrial time series data anomaly detection. During the training process of the 1D single-layer convolutional neural network model, the labels consistent with the original data are automatically assigned to the enhanced data in the time and frequency domains through the time domain aggregation feature index, and the time domain feature information is fused into the training model through this process, realizing the introduction of a time attention mechanism.

[0132] In this embodiment, during the training process of the 1D single-layer convolutional neural network model, the strategy of time-domain aggregation feature index is introduced, which not only performs a preliminary extraction of the temporal features of the data in the training data set, but also realizes automatic labeling of the enhanced data, thereby reducing the labor cost and time cost of manual labeling and improving the model training efficiency.

[0133] Figure 3 The structural block diagram of the industrial time series data anomaly detection system based on time-frequency domain feature enhancement provided by the present invention is shown in FIG. Figure 3 , the industrial time series data anomaly detection system based on time-frequency domain feature enhancement includes:

[0134] The data acquisition module 301 is used to collect industrial time series data in the industrial production process;

[0135] A time domain processing module 302, used to perform time domain processing on the industrial time series data in combination with a weighted weight adjustment strategy and a data interpolation method in the time domain;

[0136] The frequency domain enhancement module 303 is used to convert the industrial time series data after time domain processing into frequency domain data, perform frequency domain enhancement processing on the frequency domain data and then convert it back to the time domain to obtain industrial time series data enhanced in the time and frequency domains;

[0137] The anomaly detection module 304 is used to use a preset dual-branch neural network model to perform anomaly detection on the industrial time series data enhanced in the time and frequency domains.

[0138] The industrial time series data anomaly detection system based on time-frequency domain feature enhancement provided by the present invention executes the industrial time series data anomaly detection method based on time-frequency domain feature enhancement provided by the above-mentioned embodiments through the above-mentioned modules. The industrial time series data anomaly detection method based on time-frequency domain feature enhancement has been described in detail in the above-mentioned embodiments, and this embodiment will not be repeated here.

[0139] Figure 4 The structural block diagram of the electronic device provided by the present invention is as follows: Figure 4 As shown, the present invention also provides an electronic device, and the electronic device 400 can be a computing device such as a mobile terminal, a desktop computer, a notebook, a palm computer, and a server. The electronic device 400 includes a processor 401 and a memory 402, wherein the memory 402 stores an industrial time series data anomaly detection program 404 based on time-frequency domain feature enhancement.

[0140] In some embodiments, the memory 402 may be an internal storage unit of a computer device, such as a hard disk or memory of the computer device. In other embodiments, the memory 402 may also be an external storage device of the computer device, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), etc. equipped on the computer device. Furthermore, the memory 402 may also include both an internal storage unit of the computer device and an external storage device. The memory 402 is used to store application software and various types of data installed on the computer device, such as program codes for installing the computer device. The memory 402 can also be used to temporarily store data that has been output or is to be output. In one embodiment, when the industrial time series data anomaly detection program 404 based on time-frequency domain feature enhancement is executed by the processor 401, the following steps are implemented:

[0141] Collect industrial time series data in industrial production processes;

[0142] Combining the weighted weight adjustment strategy and the data interpolation method in the time domain, the industrial time series data is processed in the time domain;

[0143] Converting the industrial time series data after time domain processing into frequency domain data, performing frequency domain enhancement processing on the frequency domain data and converting it back into the time domain to obtain industrial time series data enhanced in the time and frequency domains;

[0144] A preset dual-branch neural network model is used to perform anomaly detection on industrial time series data enhanced in the time and frequency domains.

[0145] In some embodiments, the processor 401 can be a central processing unit (CPU), a microprocessor or other data processing chip, used to run the program code or process data stored in the memory 402, such as executing an industrial time series data anomaly detection program based on time-frequency domain feature enhancement.

[0146] This embodiment further provides a computer-readable storage medium, on which is stored an industrial time series data anomaly detection program based on time-frequency domain feature enhancement. When the industrial time series data anomaly detection program based on time-frequency domain feature enhancement is executed by a processor, the following steps are implemented:

[0147] Collect industrial time series data in industrial production processes;

[0148] Combining the weighted weight adjustment strategy and the data interpolation method in the time domain, the industrial time series data is processed in the time domain;

[0149] Converting the industrial time series data after time domain processing into frequency domain data, performing frequency domain enhancement processing on the frequency domain data and converting it back into the time domain to obtain industrial time series data enhanced in the time and frequency domains;

[0150] A preset dual-branch neural network model is used to perform anomaly detection on industrial time series data enhanced in the time and frequency domains.

[0151] In summary, the industrial time series data anomaly detection method and system based on time-frequency domain feature enhancement provided by the present invention have the following beneficial effects compared with the prior art:

[0152] 1) The present invention combines the weighted weight adjustment strategy and data interpolation method in the time domain to perform time domain processing on the original industrial time series data, effectively solving the information distortion problem caused by low sampling frequency and data discreteness in the data acquisition process, and providing a more reliable data basis for subsequent frequency domain enhancement and anomaly detection.

[0153] 2) The present invention converts the data after time domain processing into frequency domain and performs frequency domain enhancement processing, which can highlight the frequency domain characteristics of the data and make it easier for the model to capture abnormal signals. By converting back to the time domain through inverse fast Fourier transform, industrial time series data with enhanced time and frequency domains is obtained, which provides richer feature information for anomaly detection.

[0154] 3) The present invention adopts a preset dual-branch neural network model, which combines the advantages of convolutional neural networks in feature extraction and the expertise of long short-term memory networks in processing time series data, and can simultaneously capture the spatial and temporal characteristics of industrial time series data. This dual-branch structure enables the model to analyze data more comprehensively and improve the accuracy and efficiency of anomaly detection.

[0155] The above-described embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. 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 invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.

[0156] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting anomalies in industrial time series data based on time-frequency domain feature enhancement. It is characterized in that include: Collect industrial time series data in industrial production processes; Combining the weighted weight adjustment strategy and the data interpolation method in the time domain, the industrial time series data is processed in the time domain; Converting the industrial time series data after time domain processing into frequency domain data, performing frequency domain enhancement processing on the frequency domain data and converting it back into the time domain to obtain industrial time series data enhanced in the time and frequency domains; A preset dual-branch neural network model is used to perform anomaly detection on industrial time series data enhanced in the time and frequency domains.

2. According to claim 1, the method for detecting anomalies in industrial time series data based on time-frequency domain feature enhancement, It is characterized in that When collecting industrial time series data in an industrial production process, the method further includes: Based on Shannon sampling theorem and prior knowledge, determine the optimal sampling frequency of industrial time series data; wherein the industrial time series data includes voltage, current, temperature, humidity and speed; Determine the data capacity that needs to be expanded for industrial time series data based on the optimal sampling frequency.

3. According to claim 1, the method for detecting anomalies in industrial time series data based on time-frequency domain feature enhancement, It is characterized in that The weighted weight adjustment strategy and data interpolation method in the time domain are combined to perform time domain processing on the industrial time series data in the original data set, including: According to the characteristics of industrial time series data, a weight value is assigned to each data point of the industrial time series data; Considering the weight value of each data point, data interpolation is performed on the industrial time series data.

4. According to claim 3, the method for detecting anomalies in industrial time series data based on time-frequency domain feature enhancement, It is characterized in that After data interpolation is performed on the industrial time series data, the method further includes: The data after data interpolation is optimized; the optimization includes data rotation, data jitter, data scaling and data rearrangement.

5. According to claim 1, the method for detecting anomalies in industrial time series data based on time-frequency domain feature enhancement, It is characterized in that The industrial time series data after time domain processing is converted into frequency domain data, and the frequency domain data is subjected to frequency domain enhancement processing and then converted back to the time domain to obtain industrial time series data enhanced in the time and frequency domains, including: The industrial time series data after time domain processing is converted into frequency domain data through fast Fourier transform; Performing frequency domain enhancement processing on the frequency domain data; wherein the frequency domain enhancement processing includes frequency domain noise addition, frequency masking, frequency shifting and frequency scaling; The data that has been enhanced in the frequency domain is converted back to the time domain through inverse fast Fourier transform to obtain industrial time series data enhanced in the time and frequency domains.

6. The method for detecting anomalies in industrial time series data based on time-frequency domain feature enhancement according to claim 1, It is characterized in that The dual-branch neural network model includes a 1D single-layer convolutional neural network and a long short-term memory network; The 1D single-layer convolutional neural network and the long short-term memory network share an input layer and an output layer. A single-layer convolutional network, a feature map series connection, a maximum pooling layer, and a Softmax classification regression layer connected in sequence are arranged between the input layer and the output layer of the 1D single-layer convolutional neural network; an LSTM layer, a hidden feature layer, and a Softmax classification regression layer connected in sequence are arranged between the input layer and the output layer of the long short-term memory network.

7. The method for detecting anomalies in industrial time series data based on time-frequency domain feature enhancement according to claim 6, It is characterized in that The preset dual-branch neural network model is used to perform anomaly detection on the industrial time series data enhanced in the time and frequency domains, including: The dual-branch neural network model includes a 1D single-layer convolutional neural network and a long short-term memory network; The 1D single-layer convolutional neural network converts the industrial time series data enhanced in the time and frequency domain into two-dimensional data, slices the two-dimensional data through the single-layer convolutional network, extracts graphic features, outputs multiple feature maps, connects the feature maps in series and inputs them into the maximum pooling layer, sends the output of the maximum pooling layer to the Softmax classification regression layer, and outputs the distribution position of the abnormal data; The long short-term memory network processes the industrial time series data enhanced in the time and frequency domains through the LSTM layer to obtain the hidden features of the industrial time series data, and obtains the abnormal data prediction results through softmax classification regression; For the data to be detected, the anomaly scores are calculated respectively by 1D single-layer convolutional neural network and long short-term memory network, and the weighted sum is used to obtain the final anomaly detection result.

8. An industrial time series data anomaly detection system based on time-frequency domain feature enhancement, It is characterized in that include: Data acquisition module, used to collect industrial time series data in the industrial production process; A time domain processing module, used to perform time domain processing on the industrial time series data in combination with a weighted weight adjustment strategy and a data interpolation method in the time domain; A frequency domain enhancement module, used to convert the industrial time series data after time domain processing into frequency domain data, perform frequency domain enhancement processing on the frequency domain data and then convert it back to the time domain to obtain industrial time series data enhanced in the time and frequency domains; The anomaly detection module is used to perform anomaly detection on industrial time series data enhanced in the time and frequency domains using a preset dual-branch neural network model.

9. An electronic device, It is characterized in that comprising a memory and a processor, wherein: The memory is used to store programs; The processor is coupled to the memory and is used to execute the program stored in the memory to implement the steps in the industrial time series data anomaly detection method based on time-frequency domain feature enhancement as described in any one of claims 1 to 7 above.

10. A computer-readable storage medium, It is characterized in that Used to store computer-readable programs or instructions, which, when executed by a processor, can implement the steps in the industrial time series data anomaly detection method based on time-frequency domain feature enhancement as described in any one of claims 1 to 7 above.

Images