A dredger construction data anomaly detection method based on deep stacking strategy

Abstract

The application discloses a kind of based on deep stack strategy dredger construction data anomaly detection method, including the steps are: data preprocessing, noise and isolated point in dredger construction data are handled;Convolutional auto-encoder and CNN-LSTM neural network are constructed and trained;Convolutional auto-encoder and CNN-LSTM neural network are tested;Stacking is carried out to test data deviation, and stacking model deviation sample is established;Threshold setting is carried out using Gaussian anomaly detection algorithm to deep stack strategy model;Evaluation is carried out using evaluation index.The method of the application is simple, can effectively solve the problem that dredger construction monitoring data anomaly detection model is poor in robustness, low in recognition accuracy, and has good application prospect.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of dredger construction data anomaly detection, and particularly relates to a dredger construction data anomaly detection method based on a deep stacking strategy. BACKGROUND

[0002] The dredger operation environment is complex and changeable, and the construction monitoring data is interfered by environmental noise, equipment failure and the like, which is not conducive to the optimization design of the construction process. Establishing an accurate dredger construction data anomaly detection method is an important part of optimizing the dredger construction efficiency, and has great significance for improving the automation and intelligent level of dredging ships.

[0003] The traditional data anomaly detection method based on machine learning focuses on the outliers in the data, and it is difficult to accurately identify the group anomalies in the data. In recent years, more and more researchers have applied deep learning algorithms to the field of data mining, and have proposed various deep learning anomaly detection models, which can effectively identify anomalies in data. For example, the invention patent CN108334907A uses an autoencoder and a Gaussian model to detect anomalies in aircraft engine data, effectively extracts key features in the data, and completes data anomaly detection according to the Gaussian model. However, this method does not consider the problems of unclear feature association, multi-source data heterogeneity, parameter feature time lag and the like in the data, resulting in insufficient group anomaly recognition ability of the method for time series, which easily leads to problems such as low accuracy of anomaly detection results, poor robustness, missed detection and false detection of abnormal data and the like. SUMMARY

[0004] The purpose of the present application is to solve the problem of insufficient dredger anomaly detection capability, and a dredger construction data anomaly detection method based on a deep stacking strategy is proposed.

[0005] The dredger anomaly detection model established by the present application can reduce the problems of multi-source heterogeneity and time lag caused by dredging data, accurately identify the group anomaly position in the dredging data, solve the problems of data sample missed detection and false detection caused by threshold setting in the dredging data anomaly detection, improve the anomaly detection accuracy, and solve the problem of interference of abnormal data in the dredger construction decision-making process.

[0006] In order to achieve the above purpose, the technical scheme adopted by the present application to solve the above technical problems is:

[0007] A dredger construction data anomaly detection method based on a deep stacking strategy, comprising the following steps:

[0008] S1, data preprocessing: filtering and normalizing the dredger construction data, and dividing the processed data samples into a training set and a test set;

[0009] S2, constructing a convolutional autoencoder and a CNN-LSTM neural network using the training set data: the convolutional autoencoder is used to learn the data feature distribution in the training set, and the CNN-LSTM neural network is used to predict the data trend in the training set;

[0010] S3, precision testing of the convolutional autoencoder and the CNN-LSTM neural network using the test set data;

[0011] S4, calculating the deviation between the optimal precision test result of the convolutional autoencoder and the CNN-LSTM neural network and the true value of the test set data, and stacking the deviation to establish a stacked deviation data sample;

[0012] S5, threshold setting of the stacked deviation data sample using a Gaussian anomaly detection algorithm, and anomaly detection of the test set data sample;

[0013] S6, selecting evaluation indexes to quantitatively evaluate the anomaly detection accuracy.

[0014] Further preferred scheme, the specific content, method and step in S1 include the following:

[0015] S11, the control parameters and observations contained in the dredger construction data are screened to obtain 11 control parameters and observations that play a leading role in the dredging operation process, reduce the feature dimension, reduce the training time of the model, and improve the accuracy of the model;

[0016] S12, according to the characteristic parameters of the control parameters and observations, Hampel filtering algorithm is selected for data filtering to remove outliers, eliminate environmental noise and sensor vibration interference on the accuracy of the monitoring data, and then through normalization processing to eliminate the dimension, finally the data is divided into training set and test set two parts.

[0017] Further preferred scheme, the specific content, method and step in S2 include the following:

[0018] S21, the convolutional autoencoder includes convolutional layer, pooling layer, de-pooling layer and other structures, so that the input and output structures of the convolutional autoencoder network present obvious mirror distribution characteristics; the CNN-LSTM neural network is composed of convolutional neural network (CNN) and long short-term memory network (LSTM), wherein the CNN neural network is responsible for extracting the input data features and importing them into the LSTM neural network, realizing the prediction of the next time feature parameters by the CNN-LSTM model;

[0019] S22, the convolutional autoencoder is composed of two parts of encoder and decoder, the encoder is responsible for mapping the input multi-dimensional time series data to multiple latent spaces, and the decoder is responsible for recovering the data from the latent space, so as to realize the information reconstruction between the data and detect the time when the data does not conform to the source construction data structure. The convolutional autoencoder calculation formula is:

[0020]

[0021] Wherein, f and g are encoding and decoding processes respectively, x is the network input of the encoder, w1 and w2 are weight matrices from the input layer to the hidden layer and from the hidden layer to the output layer respectively, b1 and b2 are bias terms, and x' is the output of the autoencoder;

[0022] S23, the CNN connects an LSTM network after two convolutional operations to output the data trend prediction result. The calculation methods of the forget gate, update gate and output gate in the LSTM are as follows:

[0023] f t =σ(W xf *x t +W hf h t-1 +b f )

[0024] i t =σ(W xi *x t +W hi h t-1 +b i )

[0025] c t =f t *c t-1 +i t *tanh(W xg *x t +W hg h t-1 +b g )

[0026] o t =σ(W xo *x t +W ho h t-1 +b o )

[0027] E t =o t *tanh(c t )

[0028] Wherein, ft is the output of the forget gate, σ is the activation function of the network, W hf , W hi , W xg , W hg , W xo , W ho are the parameters of the network, i t is the output of the update gate, h t-1 is the hidden state, c t , c t-1 is the cell state, o t is the output gate, b i , b o , b g is the bias term.

[0029] S24, the convolutional autoencoder and CNN-LSTM neural network are modeled using the training set data, wherein the convolutional autoencoder has the same number of nodes as the input and output, and can identify the position of the abnormal distribution in the training set data; the CNN-LSTM neural network can predict the control parameter and observation data of the next time based on the control parameter and observation data of the current time, and can identify the abnormal time point or time period in the training set data;

[0030] Further preferred scheme, the specific content, method and steps in S3 include the following:

[0031] The precision of the convolutional autoencoder and CNN-LSTM neural network is tested based on the test set data, and the minimum absolute error is used as the index of precision. The structure of the convolutional autoencoder and CNN-LSTM neural network is adjusted to obtain the network structure with optimal precision and the test result with optimal precision;

[0032] Further preferred scheme, the specific content, method and steps in S4 include the following:

[0033] The deviation between the optimal test result of the convolutional autoencoder and CNN-LSTM neural network and the true value is calculated, and the deviation is stacked to obtain a stacked deviation data sample; wherein the stacked deviation data sample calculation method is as follows:

[0034] e=y-y'

[0035] e'=y-y AE

[0036] E=(e;e')

[0037] Wherein, e is the deviation of the test result of the CNN-LSTM neural network based on the test set data from the true data, the size is n*m; e' is the deviation of the test result of the convolutional autoencoder based on the test set data from the true data, the size is n*m; E is the stacked deviation data sample, the size is 2n*m; y is the true data of the test set, y' is the test result of the CNN-LSTM neural network based on the test set data, y AE is the test result of the convolutional autoencoder based on the test set data.

[0038] Further preferred, the specific content, method and steps in S5 include the following:

[0039] The stacked deviation data sample E is verified by a Gaussian model, the stacked deviation data sample is converted into a Gaussian distribution, the mathematical expectation and variance of the stacked deviation data sample in each dimension are obtained through distribution analysis of the sample, the probability density distribution function P of the stacked deviation data sample is calculated, and a threshold is determined based on a small amount of CrossValidation set of the test set data. When a new point is given, the probability density P calculated on the Gaussian distribution and the threshold ε calculated are used to judge whether the probability density satisfies p<ε. If yes, it is determined to be abnormal, and if not, it is determined to be normal. In this process, the mean and variance of the model are calculated first

[0040]

[0041]

[0042] At the same time, the mean and variance are used to calculate the probability density distribution function P(x)

[0043]

[0044] Wherein, m is the total number of stacked error data samples, and n is the dimension of the stacked error data sample.

[0045] Further preferred, the specific content, method and steps in S6 include the following:

[0046] The accuracy, recall rate and correct rate are used as quantitative evaluation indexes of the anomaly detection model. When the optimal results of the accuracy and the recall rate are inconsistent, the correct rate is used as the only index for model performance verification. The three evaluation indexes are calculated according to the following formulas respectively:

[0047] Accuracy (p):

[0048]

[0049] Recall rate (R):

[0050]

[0051] Accuracy:

[0052]

[0053] Among them, TP is the number of instances that are actually normal and detected as normal by the algorithm; FP is the number of instances that are actually abnormal but detected as normal by the algorithm; FN is the number of instances that are actually normal but detected as abnormal by the algorithm; TN is the number of instances that are actually abnormal and detected as abnormal by the algorithm.

[0054] The dredging ship construction data anomaly detection method based on the deep stacking strategy of the present invention has the following beneficial effects: the implementation of the detection method of the present invention can reduce the problems of multi-source heterogeneity and time lag brought by dredging data, can accurately identify the group abnormal positions in dredging data, can solve the problems of missed detection and false detection of data samples caused by threshold setting in dredging data anomaly detection, improve the accuracy of anomaly detection, and solve the problem of interference from abnormal data in the dredging ship construction decision-making process. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flowchart of the method of the present invention;

[0056] Figure 2 Figure 2 shows the Hampel filtering results of some control parameters of a trailing suction hopper dredger operating in the Yangtze River Estuary in 2018. Figure a shows the main engine speed, Figure b shows the drag head depth, and Figure c shows the drag head angle.

[0057] Figure 3 This is the anomaly detection network structure constructed by the present invention based on the deep stacking strategy. DETAILED DESCRIPTION

[0058] To further clarify the objectives, technical solutions, and advantages of the present invention, the following, in conjunction with the accompanying drawings, provides a clear and complete description of the technical solutions in the embodiments, using the 2018 construction data of a certain type of trailing suction hopper dredger at the Yangtze River Estuary as an example. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are intended to fall within the scope of protection of the present invention.

[0059] The present invention aims to use a convolutional autoencoder and a CNN-LSTM neural network stack to form a dredging ship construction data anomaly detection model based on a deep stacking strategy. This model can identify outliers in the data according to the characteristics of the dredger's historical data, improve the accuracy of anomaly detection, and provide relevant technical support for dredging operations.

[0060] like Figure 1 As shown, the present invention provides a method for detecting anomalies in dredger construction data based on a deep stacking strategy, comprising the following steps:

[0061] S1, data preprocessing: the dredger construction data is filtered and normalized, and the processed data samples are divided into training set and test set two parts; the specific content, method and step include:

[0062] S11, data preprocessing: the selected dredger construction data control parameters and observations are: engine speed, rake head depth, rake head angle, wave compensator pressure, high pressure water pump speed, mud pump speed, mud pump density, mud pump flow rate, mud pump suction vacuum, speed, overflow cylinder height. Hampel filtering is performed on the data to eliminate a large number of isolated points introduced in the construction monitoring data due to the influence of the construction environment, and to improve the uniformity of the data samples. Taking engine speed, rake head depth, and rake head angle as examples, the results are shown in Figure 2 , through Hampel algorithm, the isolated points in the data can be clearly seen, wherein, figure a engine speed data is relatively stable, there are a small amount of isolated points, figure b rake head depth data is affected by the external environment, there are a large number of isolated points in the data, figure c rake head angle data has obvious isolated points, which seriously affects the data uniformity. The isolated point detection results are shown in Table 1.

[0063] Table 1 isolated point detection results

[0064] Parameter name Parameter range Number of outliers Host speed 740-760 (r / min) 130 Rake head depth 7～23(m) 1387 Rake head angle to ground -32～0(°) 1429 Wave compensator pressure 20-32 (bar) 899 High pressure wash pump speed 48-390 (r / min) 2178 Mud pump speed 40-220 (r / min) 1861 Mud pump density 1.03 to 1.6 (kg / m 3 )]]> 517 Mud pump flow rate 0.18- (m / s) 1018 Mud pump vacuum -0.80- (bar) 1232 Speed of advance 0.5-4.7 (kn) 1578 Height of overflow barrel 8～13(m) 282

[0065] S12, normalize the data, and divide the data into training set and test set according to the ratio of 4:1.

[0066] S2, use the training set data to construct a convolutional autoencoder and a CNN-LSTM neural network: use the convolutional autoencoder to learn the data feature distribution in the training set, and use the CNN-LSTM neural network to predict the data trend in the training set; the specific content, method and step include:

[0067] S21, construct a convolutional autoencoder and a CNN-LSTM neural network and generate a deep stacking strategy anomaly detection model as shown in Figure 3 : use Matlab to build a convolutional autoencoder, which consists of an input layer, two convolutional layers, a pooling layer, an inverse pooling layer, and an output layer. After construction, the initial hyperparameters are set, and the filtered and normalized training set data is imported into the convolutional autoencoder. At the same time, a CNN-LSTM network is built, which consists of two layers of CNN and one layer of LSTM. After construction, the initial hyperparameters are set, and the Hampel filtered and normalized training set data is imported into the CNN-LSTM neural network.

[0068] S22, based on the CNN neural network, a convolutional autoencoder is established, which consists of an encoder and a decoder. The encoder is responsible for mapping the input multi-dimensional time series data to multiple latent spaces, while the decoder is responsible for recovering data from the latent space, thereby realizing information reconstruction between data and detecting the time when the data does not conform to the source data structure. The formula for calculating the autoencoder is:

[0069]

[0070] Where f and g are the encoding and decoding processes, respectively, x is the network input of the encoder, w1 and w2 are the weight matrices from the input layer to the hidden layer and from the hidden layer to the output layer, respectively, b1 and b2 are the bias terms, and x' is the output of the autoencoder.

[0071] S23, the CNN connects an LSTM network after two convolution operations to output the data trend prediction result. The calculation methods of the forget gate, update gate and output gate in the LSTM are as follows:

[0072] f t =σ(W xf *x t +W hf h t-1 +b f )

[0073] i t =σ(W xi *x t +W hi h t-1 +b i )

[0074] c t =f t *c t-1 +i t *tanh(W xg *x t +W hg h t-1 +b g )

[0075] o t =σ(W xo *x t +W ho h t-1 +b o )

[0076] E t =o t *tanh(c t )

[0077] where f t is the forget gate output, σ is the activation function of the network, W hf , W hi , W xg , W hg , W xo , W ho are the parameters of the network, i t is the update gate output, h t-1 is the hidden state, c t , c t-1 is the cell state, o t is the output gate, b i , b o , b g are the bias terms.

[0078] S24, the convolutional autoencoder and the CNN-LSTM model take the host speed, the rake head depth, the rake head angle, the wave compensator pressure, the high-pressure flush pump speed, the mud pump speed, the mud pump density, the mud pump flow rate, the mud pump suction vacuum, the speed, and the overflow cylinder height as inputs, wherein the convolutional autoencoder has the same number of nodes for input and output, and can identify the position of the abnormal distribution in the dredging data; the CNN-LSTM neural network can predict the control parameters and observation data at the next moment based on the control parameters and observation data at the current moment, and can identify the time point or time period of the abnormal dredging data in the data, and the training hyperparameter settings of the convolutional autoencoder and the CNN-LSTM neural network are shown in Table 2.

[0079] Table 2 Training hyperparameter settings of the convolutional autoencoder and the CNN-LSTM neural network

[0080] Parameter name Parameter meaning Parameter value Optimizer Optimizer Adam Learning rata Learning rate 0.001 Batch size Batch training sample number 1000 Epochs Training rounds 100

[0081] S3, the precision of the convolutional autoencoder and the CNN-LSTM neural network is tested based on the test set data, the minimum absolute error is taken as the index of the precision, the structure of the convolutional autoencoder and the CNN-LSTM neural network is adjusted, the network structure with the optimal precision and the test result with the optimal precision are obtained;

[0082] S4, the deviation between the test result with the optimal precision of the convolutional autoencoder and the CNN-LSTM neural network and the true value is calculated, and the deviation is stacked to obtain a stacked deviation data sample; the specific content, method and steps are as follows:

[0083] e = y - y'

[0084] e' = y - y AE

[0085] E = (e; e')

[0086] Wherein, e is the deviation of the test result of the CNN-LSTM neural network based on the test set data from the true data, the size is n*m; e' is the deviation of the test result of the convolutional autoencoder based on the test set data from the true data, the size is n*m; E is the stacked deviation data sample, the size is 2n*m; y is the true data of the test set, y' is the test result of the CNN-LSTM neural network based on the test set data, y AE is the test result of the convolutional autoencoder based on the test set data.

[0087] S5, the specific content, method and steps of setting threshold value and performing abnormal detection on the stacked error data sample by using Gaussian anomaly detection algorithm are as follows: Gaussian model verification is performed on the model error data sample, the data is converted into Gaussian distribution, the mathematical expectation and variance of the data in each dimension are obtained through distribution analysis of the sample, the probability density distribution function of the data is calculated, and a threshold value is selected based on a small amount of Cross Validation set of the test set data. When a new point is given, the probability density P calculated on the Gaussian distribution and the threshold value calculated are used to judge whether the probability density satisfies p<ε. If yes, it is determined as abnormal, and if not, it is determined as normal. In this process, the mean and variance of the model are calculated as follows:

[0088]

[0089]

[0090] At the same time, the mean and variance are used to calculate the probability density distribution function P(x)

[0091]

[0092] Wherein, m is the total number of stacked error data samples, and n is the dimension of the stacked error data sample.

[0093] S6, the specific content, method and steps of evaluating the deep stacked anomaly detection model formed by the convolutional autoencoder and the CNN-LSTM neural network are as follows: after classifying the results, the performance of the algorithm is evaluated by calculating the accuracy, recall rate and correct rate, and the calculation method is as follows:

[0094] Accuracy (p):

[0095]

[0096] Recall rate (R):

[0097]

[0098] Accuracy (correct rate):

[0099]

[0100] Wherein, TP is the actual number of normal instances and the algorithm detects normal; FP is the actual number of abnormal instances but the algorithm detects normal; FN is the actual number of normal instances but the algorithm detects abnormal; TN is the actual number of abnormal instances and the algorithm detects abnormal instances.

[0101] The abnormal detection results of the convolutional autoencoder AE, the CNN-LSTM neural network and the deep stack strategy anomaly detection method ACLG are shown in Table 3.

[0102] Table 3 Model comparison results

[0103] Model Accuracy (p) Recall rate (R) Accuracy CNN-LSTM 0.9674 0.9896 0.9607 AE 0.9359 0.986 0.9252 ACLG 0.9870 0.9896 0.9771

[0104] Through the analysis of the above Table 3, it can be seen that the deep stack strategy anomaly detection method ACLG used in the application has the best accuracy, recall rate and correct rate in the test set data among the three models, indicating that the deep stack strategy anomaly detection method ACLG can effectively identify the group anomaly in the data and improve the accuracy of anomaly detection. Therefore, the deep stack strategy anomaly detection method is the best choice for anomaly detection in the dredging operation of the dredger.

[0105] From the above experimental results, it can be seen that the dredger construction data anomaly detection method based on the deep stack strategy is more excellent as a whole, the performance is relatively stable, can effectively reduce the data imbalance caused by the abnormal value in the data, reduce the multi-source heterogeneous and time delay problems caused by the dredging data, can accurately identify the group anomaly position in the dredging data, improve the accuracy of anomaly detection, solve the problem of interference by abnormal data in the dredger construction decision-making process. It has better application prospect in the dredging anomaly detection of the drag suction dredger, and according to the analysis of the construction environment and data distribution characteristics of the drag suction dredger, the application can also be applied in the dredging operation field of the drag suction dredger.

Claims

1. A method for detecting abnormality of dredger construction data based on deep stacking strategy, characterized in that, It comprises the following steps: S1, data preprocessing: filtering and normalizing the dredger construction data, and dividing the processed data samples into training set and test set two parts; S2, using training set data to construct convolutional autoencoder and CNN-LSTM neural network: using convolutional autoencoder to learn the data feature distribution in the training set, and using CNN-LSTM neural network to predict the data trend in the training set; S3, using test set data to test the precision of convolutional autoencoder and CNN-LSTM neural network; S4, calculate the deviation between the optimal precision test result of convolutional autoencoder and CNN-LSTM neural network and the true value of test set data, and stack the deviation to establish the stacked deviation data sample; S5, using Gaussian anomaly detection algorithm to set threshold for stacked deviation data sample, and testing the anomaly detection of test set data sample; S6, select evaluation index to quantitatively evaluate the anomaly detection precision.

2. The method according to claim 1, wherein the method is characterized by: The specific content and method of data preprocessing in step S1 include the following steps: S11, select the control parameters and observations contained in the dredger construction data in multiple construction periods, and obtain the control parameters and observations that play a leading role in the dredging operation process; S12, according to the data characteristics of the selected control parameters and observations, use Hampel filtering and normalization method to preprocess the dredger construction data, eliminate environmental noise and the problem of different data dimensions of sensor data.

3. The method according to claim 1, wherein the method is characterized by: Step S2, the specific content and method of constructing convolutional autoencoder and CNN-LSTM neural network include the following steps: S21, the convolutional autoencoder includes convolutional layer, pooling layer and inverse pooling layer structure, so that the input and output structure of the convolutional autoencoder network presents obvious mirror distribution; the CNN-LSTM neural network is composed of convolutional neural network CNN and long short term memory network LSTM, wherein the convolutional neural network CNN is responsible for extracting the input data features and importing them into the LSTM neural network, realizing the prediction of the next time control parameters and observations by CNN-LSTM neural network; S22, the calculation formula of convolutional autoencoder is: Where f and g are encoding and decoding process respectively, x is the network input of encoder, w1 and w2 are weight matrix from input layer to hidden layer and from hidden layer to output layer respectively, b1 and b2 are bias terms, and x' is the output of autoencoder; S23, after two layers of convolution operation, CNN is connected with an LSTM neural network to realize data trend prediction, wherein the calculation method of forget gate, update gate and output gate in LSTM network is as follows: f t = σ(W xf *x t+Whfht-1+bf ) i t = σ(W xi*xt+Whiht-1+bi ) c t = f t*ct-1+it*tanh (W xg*xt+Whght-1+bg ) o t = σ(W xo*xt+Whoht-1+bo ) E t = o t tanh(c t ) wherein f t is the forget gate output, σ is the activation function of the network, W hf , W hi , W xg , W hg , W xo , W ho are parameters of the network, i t is the update gate output, h t-1 is the hidden state, c t , c t-1 is the cell state, o t is the output gate, b i , b o , b g are bias terms; S24, training the convolutional autoencoder and the CNN-LSTM neural network using the training set data, wherein the convolutional autoencoder has the same number of nodes for input and output, and is capable of identifying the location of abnormal distribution in the training set data; the CNN-LSTM neural network is capable of identifying abnormal time points or time periods in the training set data by predicting the control parameters and observation data at the next time point based on the control parameters and observation data at the current time.

4. The method according to claim 1, wherein the method is characterized by: Step S3, the specific method of using test set data to test the precision of convolutional autoencoder and CNN-LSTM neural network is to use the minimum mean absolute error as the precision index to adjust the structure of convolutional autoencoder and CNN-LSTM neural network, and obtain the optimal test set data test result.

5. The method according to claim 1, wherein the method is characterized by: Step S4, the deviation between the test results of the convolutional auto-encoding and CNN-LSTM neural network and the true values of the test set data is calculated, and the deviation is stacked to establish a specific method and calculation formula for the stacked deviation data sample as follows: e=y-y' e' = y - y AE E=(e;e') Wherein, e is the deviation of the test result of the CNN-LSTM neural network based on the test set data from the real data, the size is n*m; e' is the deviation of the test result of the convolutional autoencoder based on the test set data from the real data, the size is n*m; E is the stacked deviation data sample, the size is 2n*m; y is the real data of the test set, y' is the test result of the CNN-LSTM neural network based on the test set data, y AE is the test result of the convolutional autoencoder based on the test set data.

6. The method according to claim 1, wherein the method is characterized by: Step S5, the specific content and method steps of the Gaussian anomaly detection algorithm for threshold setting of the stacked deviation data sample and the abnormality detection of the test set data sample include: The stacked deviation data sample is converted into a Gaussian distribution, the distribution of the sample is analyzed, the mathematical expectation and variance of the stacked deviation data sample in each dimension are obtained, the probability density distribution function of the stacked deviation data sample is calculated, and a threshold is determined based on a small amount of Cross Validation set of the stacked deviation data sample; when a new point is given, the probability density P calculated on the Gaussian distribution and the threshold ε calculated are used to judge whether the current probability density meets p<ε, if yes, it is determined as abnormal, otherwise, it is determined as normal, in this process, the mean and variance of the stacked deviation data sample are calculated as follows: At the same time, the mean and variance are used to calculate the probability density distribution function P(x) Where m is the total number of stacked error data samples, and n is the dimension of the stacked error data sample.

7. The method according to claim 1, wherein the method is characterized by: Step S6, the specific content of the quantitative evaluation of the abnormality detection accuracy includes accuracy, recall rate, and correct rate, when the optimal results of the accuracy and the recall rate are inconsistent, the correct rate is used as the only index for model performance verification, and the respective index definitions and calculation formulas are as follows: Accuracy p: Recall rate R: Accuracy correct rate: Where TP is the number of instances that are actually normal and detected as normal by the algorithm; FP is the number of instances that are actually abnormal but detected as normal by the algorithm; FN is the number of instances that are actually normal but detected as abnormal by the algorithm; and TN is the number of instances that are actually abnormal and detected as abnormal by the algorithm.

Images