CN116108885A - An abnormal detection method for medium and low pressure gas pressure regulators based on LSTM-AE-DT model - Google Patents

Abstract

The present invention relates to a method for detecting abnormalities of medium and low pressure gas pressure regulators based on LSTM-AE-DT model, comprising: S1, collecting outlet pressure data of medium and low pressure gas pressure regulators; S2, pre-setting the outlet pressure data of gas pressure regulators Processing; S3, establishing the LSTM‑AE model, inputting the preprocessed training data to train the model, and saving the trained model; S4, inputting the preprocessed test data into the trained algorithm model, and outputting the gas adjustment Predicted value of the outlet pressure of the pressure regulator; S5. Calculate the error and threshold ε between the predicted value of the outlet pressure of the gas pressure regulator and the actual value of the outlet pressure of the gas pressure regulator; S6. Detect the state of the time series data of the gas pressure regulator according to the threshold ε , output the corresponding operating status of the gas pressure regulator. Compared with the prior art, the present invention combines the long-term short-term memory network and the self-encoder, can reconstruct the time-series data on the basis of learning the long-term dependence of the data, and improves the abnormal detection performance of the time-series data of the gas pressure regulator .

Description

A method for detecting abnormality of medium and low pressure gas regulators based on LSTM-AE-DT model

Technical Field

The present invention relates to a method for detecting anomalies of a medium- and low-pressure gas pressure regulator, and in particular to a method for detecting anomalies of a medium- and low-pressure gas pressure regulator based on an LSTM-AE-DT model.

Background Art

With the rapid development of the gas industry, how to ensure the stability and safety of gas transportation has always been the focus of research in the gas field. The gas pressure regulator is an important equipment to ensure the normal transportation of gas to downstream users. When it fails, it will cause problems such as insufficient downstream gas consumption and unstable gas supply. If the fault is not discovered and repaired in time, it will continue to deteriorate and cause accidents such as explosions and fires. The time series data during the operation of the gas pressure regulator can be used to feedback the status of the pressure regulator, and then detect whether the gas pressure regulator is faulty. Therefore, the research on anomaly detection of gas pressure regulator time series data has great practical significance and value.

Gas pressure regulator anomaly detection technology is mainly used to determine whether the gas pressure regulator is in normal working state or abnormal state. The traditional method of gas pressure regulator anomaly detection technology is to directly judge the anomaly through the experience of maintenance workers or use traditional statistical theory to judge the anomaly; the traditional method cannot warn of potential faults in advance, and has low accuracy and poor timeliness. Therefore, the research on intelligent detection methods for anomaly detection of gas pressure regulator time series data is an urgent problem to be solved.

In recent years, there are more and more machine learning methods and their applications are becoming more and more extensive. Many scholars in the industry have proposed gas

Gas pressure regulator anomaly detection technology. The machine learning-based gas pressure regulator anomaly detection technology uses the system's historical data under normal and fault conditions to train support vector machines and other machine learning algorithms for anomaly detection; this technology requires manual labeling of data for supervised processing, which consumes a lot of time and labor costs.

Summary of the invention

The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a method for detecting abnormalities of medium and low pressure gas pressure regulators based on the LSTM-AE-DT model.

The purpose of the present invention can be achieved by the following technical solutions:

A method for detecting abnormalities of medium and low pressure gas pressure regulators based on LSTM (Long Short-Term Memory Network)-AE (Autoencoder)-DT (Dynamic Threshold Method) model comprises the following steps:

S1. Collect outlet pressure data of medium and low pressure gas pressure regulators;

S2. Preprocess the gas pressure regulator outlet pressure data to obtain training data and test data;

S3. Establish an LSTM-AE model, input the preprocessed training data to train the model, and save the trained model;

S4, input the preprocessed test data into the trained algorithm model, and output the predicted value of the gas pressure regulator outlet pressure;

S5, the error between the predicted value of the gas pressure regulator outlet pressure output by reconstruction and the true value of the gas pressure regulator outlet pressure, and the threshold ε of this section of time series data;

S6. Detect the state of the gas pressure regulator timing data according to the threshold ε, and output the normal operation state or abnormal operation state of the gas pressure regulator.

Furthermore, in the step S2, the original outlet pressure time series training data set X=(x ₁ , x ₂ , …, x _j , …, x _N ), 1≤j≤N, of the gas pressure regulator is obtained through step S1, and the average value of the outlet pressure value at the same time point three days before the missing data is calculated to fill the missing data value; the original time series data set is processed by using the deviation standardization method to obtain a standardized time series data set X′=(x′ ₁ , x′ ₂ , …, x′ _j , …, x′ _N ), x′ _j ∈[0,1], and the formula is as follows:

Among them, _xj is the jth original data, _x′j is the jth standardized time series data, min _1≤j≤N { _xj } represents the minimum value in the training data set, and max _1≤j≤N { _xj } represents the maximum value in the training data set; finally, the data set X′ is framed by the sliding window method, the sliding window width is l, 1≤l≤N, and the sliding step is set to 1. At this time, the data set after the frame operation is Y=( _y1 , _y2 ,…, _yN-l+1 ), where _yi =( _x′i , _x′i+1 ,…, _x′i+l-1 ), 1≤i≤N-l+1.

Furthermore, in step S3, the LSTM network uses a forget gate to filter information, the input gate retains necessary information and encodes it, and the hidden state is output through the output gate and passed to the next LSTM unit for training to extract features from the data.

Furthermore, the step S3 constructs an LSTM network on the encoder and decoder of AE, and the encoder obtains a high-dimensional input data sequence in the form of a vector of a fixed size.

Furthermore, in step S4, the framed data _yi = ( _x'i , x'i ₊₁ , ..., _x'i+l-1 ) is input into the LSTM-AE model for training to obtain a reconstructed time series data set.

Furthermore, in step S5, the reconstruction error e _j between the predicted value of the gas pressure regulator outlet pressure and the true value of the gas pressure regulator outlet pressure is calculated as follows:

表示x′_j第k次重构的出口压力数据，由x′_j的重构误差e_j的计算方法得到重构误差序列：e＝(e₁，e₂，…，e_j，…，e_N)，对e进行指数加权滑动平均，得到误差序列e′＝(e′₁，e′₂，…，e′_j，…，e′_N)，之后计算阈值序列ε：Where x′ _j represents the jth outlet pressure data, n represents the number of times x′ _j is reconstructed,

represents the outlet pressure data reconstructed for the kth time by x′ _j . The reconstruction error sequence is obtained by calculating the reconstruction error e _j of x′ _j : e = (e ₁ , e ₂ , …, e _j , …, e _N ). The error sequence e′ = (e′ ₁ , e′ ₂ , …, e′ _j , …, e′ _N ) is obtained by performing exponential weighted sliding average on e. Then, the threshold sequence ε is calculated:

ε＝μ(e′)+Z·σ(e′)

计算f(ε_t)：Among them, μ(·) is the mean, σ(·) is the standard deviation, and Z represents the artificially set weight.

Calculate f(ε _t ):

Δμ(e′)=μ(e′)-μ({e∈e′|e<ε _t })

Δσ(e′)=σ({e∈e′|e<ε _t })

e _a ={e∈e′|e＞ε _t }

Wherein, E _seq is a set of continuous data in _ea , the threshold ε that makes f(ε _t ) reach the maximum value is the required threshold, and anomaly detection of time series data is performed based on ε.

Furthermore, in step S5, the LSTM-AE-DT model is optimized by using a back propagation algorithm, and the network parameters are updated until the input and output loss functions L reach a minimum value. The formula of the loss function L is:

Wherein, _ej is the reconstruction error between the predicted value of the gas pressure regulator outlet pressure and the true value of the gas pressure regulator outlet pressure.

Furthermore, in step S6, precision, recall and F1 value are used as evaluation indicators, and the calculation formula is:

Among them, Precision is the precision, Recall is the recall, TP represents the number of abnormal data predicted as abnormal data; FP represents the number of normal data predicted as abnormal data; FN represents the number of abnormal data predicted as normal data.

Furthermore, the LSTM-AE model in step S3 utilizes the storage capacity of LSTM so that the data processed by the encoder maintains the long-term dependency of the time series data, while compressing the high-dimensional input vector into a low-dimensional vector.

Furthermore, in step S5, a dynamic threshold method is used to calculate the threshold ε of the time series data.

Compared with the prior art, the present invention has the following beneficial effects:

1. The present invention combines a long short-term memory network with an autoencoder. The long short-term memory network is composed of multiple LSTM units. The autoencoder can better extract features from non-stationary time series data and eliminate the influence of noise on time series data. The model established by combining the two adopts an unsupervised algorithm. During the training process, it is not necessary to label the data. The long-term dependence of the data in the time series is learned and the correlation between the data is extracted, thereby reconstructing the time series data of the gas pressure regulator, saving time and reducing labor costs.

2. The present invention uses a dynamic threshold method, fully considering the long memory in the time series data, avoiding information loss caused by too long a time series sequence, dynamically adjusting the threshold according to the changes in the time series data, realizing automatic updating of the threshold for anomaly detection, and improving the performance of anomaly detection of the gas pressure regulator time series data.

BRIEF DESCRIPTION OF THE DRAWINGS

Fig. 1 is a schematic diagram of a process of the present invention;

FIG2 is a schematic diagram of the LSTM-AE model structure of the present invention;

FIG. 3 is a schematic diagram of a curve showing abnormality detection results of a medium- and low-pressure gas pressure regulator according to the present invention.

DETAILED DESCRIPTION

The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

As shown in FIG1 , a method for detecting abnormalities of medium and low pressure gas pressure regulators based on the LSTM-AE-DT model is shown, which includes the following steps:

S1. Collect outlet pressure data of medium and low pressure gas pressure regulators;

S2. Preprocess the gas pressure regulator outlet pressure data to obtain training data and test data;

S3. Establish an LSTM-AE model, input the preprocessed training data to train the model, and save the trained model;

S4, input the preprocessed test data into the trained algorithm model, and output the predicted value of the gas pressure regulator outlet pressure;

S5, the error between the predicted value of the gas pressure regulator outlet pressure output by reconstruction and the true value of the gas pressure regulator outlet pressure, and the threshold ε of this section of time series data;

S6. Detect the state of the gas pressure regulator timing data according to the threshold ε, and output the normal operation state or abnormal operation state of the gas pressure regulator.

In step S2, the original outlet pressure time series training data set X=( _x1 , _x2 , ..., _xj , ..., _xN ), 1≤j≤N of the gas pressure regulator is obtained through step S1, and the average value of the outlet pressure value at the same time point three days before the missing data is calculated to fill the missing data value; the original time series data set is processed by using the deviation standardization method to obtain a standardized time series data set X′=( _x′1 , _x′2 , ..., _x′j , ..., _x′N ), _x′j∈ [0,1], and the formula is as follows:

Among them, _xj is the jth original data, _x′j is the jth standardized time series data, min _1≤j≤N { _xj } represents the minimum value in the training data set, and max _1≤j≤N { _xj } represents the maximum value in the training data set; finally, the data set X′ is framed by the sliding window method, the sliding window width is l, 1≤l≤N, and the sliding step is set to 1. At this time, the data set after the frame operation is Y=( _y1 , _y2 ,…, _yN-l+1 ), where _yi =( _x′i , _x′i+1 ,…, _x′i+l-1 ), 1≤i≤N-l+1.

In step S3, the LSTM network uses the forget gate to filter information, the input gate retains necessary information and encodes it, and the hidden state is output through the output gate and passed to the next LSTM unit for training to extract features from the data. The LSTM network is built on the encoder and decoder of AE, and the encoder obtains high-dimensional input data sequences in the form of fixed-size vectors.

In step S4, the framed data _yi = ( _x'i , x'i ₊₁ , ..., _x'i+l-1 ) is input into the LSTM-AE model for training to obtain a reconstructed time series data set.

In step S5, the reconstruction error e _j between the predicted value of the gas pressure regulator outlet pressure and the true value of the gas pressure regulator outlet pressure is calculated:

表示x′_j第k次重构的出口压力数据，由x′_j的重构误差e_j的计算方法得到重构误差序列：e＝(e₁，e₂，…，e_j，…，e_N)，对e进行指数加权滑动平均，得到误差序列e′＝(e′₁，e′₂，…，e′_j，…，e′_N)，之后计算阈值序列ε：Where x′ _j represents the jth outlet pressure data, n represents the number of times x′ _j is reconstructed,

represents the outlet pressure data reconstructed for the kth time by x′ _j . The reconstruction error sequence is obtained by calculating the reconstruction error e _j of x′ _j : e = (e ₁ , e ₂ , …, e _j , …, e _N ). The error sequence e′ = (e′ ₁ , e′ ₂ , …, e′ _j , …, e′ _N ) is obtained by performing exponential weighted sliding average on e. Then, the threshold sequence ε is calculated:

ε＝μ(e′)+Z·σ(e′)

计算f(ε_t)：Among them, μ(·) is the mean, σ(·) is the standard deviation, and Z represents the artificially set weight.

Calculate f(ε _t ):

Δμ(e′)=μ(e′)-μ({e∈e′|e<ε _t })

Δσ(e′)=σ({e∈e′|e<ε _t })

e _a ={e∈e′|e＞ε _t }

Wherein, E _seq is a set of continuous data in ea , the threshold ε that makes f(ε _t ) reach the maximum value is the required threshold, and anomaly detection of time series data is performed based on ε.

The LSTM-AE-DT model is optimized by using the back propagation algorithm, and the network parameters are updated until the input and output loss function L reaches the minimum value. The formula of the loss function L is:

Wherein, _ej is the reconstruction error between the predicted value of the gas pressure regulator outlet pressure and the true value of the gas pressure regulator outlet pressure.

In step S6, precision, recall and F1 value are used as evaluation indicators, and the calculation formula is:

Among them, Precision is the precision, Recall is the recall, TP represents the number of abnormal data predicted as abnormal data; FP represents the number of normal data predicted as abnormal data; FN represents the number of abnormal data predicted as normal data.

In step S3, the LSTM-AE model uses the storage capacity of LSTM to ensure that the data processed by the encoder maintains the long-term dependency of time series data, while compressing the high-dimensional input vector into a low-dimensional vector.

In step S5, a dynamic threshold method is used to calculate the threshold ε of the time series data.

Theoretical basis of the present invention:

(a) Long short-term memory network model

Long short-term memory network is a variant of recurrent neural network RNN. The information of RNN hidden layer only comes from the current input and the information of the hidden layer at the previous moment, and it has no memory function. In order to solve the problem that RNN cannot achieve long-term dependence, LSTM model introduces cell state and uses three gates: input gate, forget gate and output gate to maintain and control information.

Read the hidden state h _t-1 of the output gate at the previous moment and the input x _t at the current moment, and use the Sigmoid activation function δ(·) to calculate the output _ft of the forget gate. Use _ft to control the information that needs to be forgotten in the cell state C _t-1 at the previous moment. Among them, W _f and b _f are the weight and bias of the forget gate.

f _t =δ (W _f [h _t-1 , x _t ]+b _f )

控制上一层h_t-1和x_t通过输入门需要保留的信息。其中，W_i和W_c分别为输入门和细胞状态的权重，b_i和b_c分别为输入门和细胞状态的偏置。Calculate the value of the input gate i _t and the temporary cell state at the current moment

Control the information that needs to be retained by the previous layer h _t-1 and x _t through the input gate. Among them, _Wi and _Wc are the weights of the input gate and cell state, respectively, and _bi and _bc are the biases of the input gate and cell state, respectively.

i _t = δ (W _i [h _t-1 , x _t ]+b _i )

Calculate the current cell state C _t , the value of the output gate o _t and the hidden state of the output gate h _t . Among them, W _o and b _o are the weight and bias of the output gate respectively.

o _t = δ (W _o [h _t-1 , x _t ]+b _o )

h _t = o _t tan(C _t )

再结合遗忘门的输出f_t和输入门的值i_t更新当前时刻细胞状态C_t。最后，结合输出门的值o_t将内部信息传递到外部隐藏状态h_t。The LSTM unit first uses the hidden state of the output gate at the previous moment h _t-1 and the input x _t at the current moment to calculate the temporary cell states of the three gates and the current moment.

Then combine the output of the forget gate _ft and the value of the input gate _{it to} update the current cell state _Ct . Finally, combine the value of the output gate _ot to pass the internal information to the external hidden state _ht .

(b) Autoencoder model

The autoencoder is a neural network model proposed in 1986, which consists of an encoder and a decoder. The encoder performs the encoding process of the model, and the decoder performs the decoding process of the model. The parameters of the autoencoder are optimized through the back propagation algorithm so that the loss error between the input and output reaches the minimum value. At this time, the model is trained to the optimal value.

The input data X is encoded and weighted and summed. Combined with the bias of the encoder, the encoding Y is calculated through the activation function f _m (·). Among them, R _m and b _m are the weight and bias of the encoder respectively.

Y＝f _m (R _m X+b _m )

其中，R_n和b_n分别为解码器的权重和偏置。The code Y is calculated in the same way and output through the activation function f _n (·)

Among them, _Rn and _bn are the weight and bias of the decoder respectively.

The mean square error is selected as the loss function L, and the gradient descent method is used to minimize the loss function L. The back propagation algorithm gradually updates the model parameters θ so that the loss error continues to approach the minimum value, and the minimum loss error E can be obtained at this time.

The embodiment of the present invention provides a method for detecting abnormalities of medium and low pressure gas pressure regulators based on the LSTM-AE-DT model. During data preprocessing, the sliding window width is set to 60 and the sliding step size is set to 1. When calculating the threshold using the dynamic threshold method, the weight Z=(1.25, 1.30, ... 2.0) is set according to the characteristics of the gas pressure regulator outlet pressure time series data. The present invention uses Python 3.7 to build a model based on TensorFlow 2.2, uses the Windows 11 operating system, and conducts experiments on a hardware device with an Intel i5-12500H processor and 16G memory. The specific steps are as follows:

Step 1: Collect outlet pressure data of medium and low pressure gas pressure regulators;

The specific process is as follows:

A gas regulator outlet pressure data set collected by an energy company through the SCADA system from June 25, 2021 to October 9, 2021 was selected. Data was collected every 5 minutes, and a total of 30,528 data were collected. The first 21,370 data were used as training sets, and the last 9,158 data were used as test sets.

Step 2: preprocessing the gas pressure regulator outlet pressure data;

The specific process is as follows:

Assume that the original outlet pressure time series training data set of the gas pressure regulator is X = (x ₁ , x ₂ , …, x _j , …, x _N ), 1≤j≤N. First, the missing data values are filled by calculating the average value of the outlet pressure value at the same time point three days before the missing data value in the data obtained in step 1. Then, the deviation standardization method is used to process the original time series data set obtained in step 1, and a standardized time series data set X′ = (x′ ₁ , x′ ₂ , …, x′ _j , …, x′ _N ), x′ _j ∈ [0, 1], can be obtained, and the formula is as follows:

Among them, _xj is the jth original data, _x′j is the jth standardized time series data, min _1≤j≤N { _xj } represents the minimum value in the training data set, and max _1≤j≤N { _xj } represents the maximum value in the training data set; finally, the data set X′ is framed by the sliding window method, the sliding window width is l, 1≤l≤N, and the sliding step is set to 1. At this time, the data set after the frame operation is Y=( _y1 , _y2 ,…, _yN-l+1 ), where _yi =( _x′i , _x′i+1 ,…, _x′i+l-1 ), 1≤i≤N-l+1.

Step 3: Establish an LSTM-AE model, input the preprocessed training data to train the model, and save the trained model;

The specific process is as follows:

In the LSTM-AE model, LSTM uses a forget gate to filter information, the input gate retains necessary information and encodes it, and the hidden state is output through the output gate and passed to the next LSTM unit for training, thereby realizing data feature extraction. An LSTM network is built on the encoder and decoder of AE, and the encoder obtains a high-dimensional input data sequence in the form of a vector of a fixed size. By utilizing the storage capacity of LSTM, the data processed by the encoder still maintains the long-term dependency of the time series data, while compressing the high-dimensional input vector into a low-dimensional vector. The structure of the LSTM-AE model is shown in Figure 2.

和当前时刻的输入单元x′_i。第一个单元编码得到当前时刻的隐藏状态h_t和临时细胞状态

传递到第二个单元，第二个单元决定是否保留第一个单元信息，依次向下传递，直至最后一个LSTM单元。所有数据的信息将通过最后一个LSTM单元输出，输出结果为Z，将Z进行重复编码得RV_Z1，RV_Z2，…，RV_Zl。编码过程如图2LSTM编码器所示。The main function of the LSTM encoder is to implement the encoding process of the LSTM-AE model and learn the rules of time series data features. Each time series data y _i in the preprocessed data Y is input into the LSTM unit in sequence. The input of a single LSTM unit is the hidden state h _t-1 and the temporary cell state obtained by the LSTM unit encoding at the previous moment.

and the current input unit x′ _i . The first unit encodes the current hidden state h _t and the temporary cell state

The information is passed to the second unit, which decides whether to keep the information of the first unit, and is passed down in sequence until the last LSTM unit. All data information will be output through the last LSTM unit, and the output result is Z. Z is repeatedly encoded to obtain RV _Z1 , RV _Z2 , ..., RV _Zl . The encoding process is shown in Figure 2 LSTM encoder.

The main function of the LSTM decoder is to realize the decoding process of the LSTM-AE model. The encoder passes the information in RV _Z1 , RV _Z2 , ..., RV _Zl to the decoder as the input of the LSTM unit. The decoder is then used to reconstruct the output at time (N-l+1) and calculate the hidden state at time (N-l+1), thereby realizing the reconstruction of the time series data at time (N-l+1). This is passed in sequence until the last LSTM unit is calculated. The decoding process is shown in Figure 2 LSTM decoder.

x′_j表示第j个出口压力数据，n表示x′_j重构的次数，

表示x′_j第k次重构的出口压力数据，计算燃气调压器出口压力的预测值与燃气调压器出口压力的真实值的重构误差e_j：The original data is normalized by deviation to obtain the outlet pressure data set X′＝(x′ ₁ ，x′ ₂ ，…，x′ _j ，…，x′ _N ), x′ _j ∈[0，1], which is framed and the framed data y _i ＝(x′ _i ，x′ _i+1 ，…，x′ _i+l-1 ), 1≤i≤N-l+1 is input into the LSTM-AE model. The loss function of the reconstruction error calculation model is used to train the model. The reconstructed time series data set

x′ _j represents the jth outlet pressure data, n represents the number of times x′ _j is reconstructed,

x′ _j represents the outlet pressure data reconstructed for the kth time, and the reconstruction error e _j between the predicted value of the outlet pressure of the gas pressure regulator and the true value of the outlet pressure of the gas pressure regulator is calculated:

The back propagation algorithm is used to optimize the model and update the network parameters until the loss function of the input and output is minimized.

Step 4: Input the preprocessed test data into the trained algorithm model, use a sliding window of l, input l data, and predict the l+1th data, thereby outputting the predicted value of the gas pressure regulator outlet pressure;

Step five, reconstruct the error between the predicted value of the gas pressure regulator outlet pressure and the true value of the gas pressure regulator outlet pressure, and use the dynamic threshold method to calculate the threshold of the time series data segment.

The specific process is as follows:

The reconstruction error sequence is obtained by calculating the reconstruction error e _j of x′ _j : e = (e ₁ , e ₂ , …, e _j , …, e _N ), and an exponentially weighted sliding average is performed on e to obtain the error sequence e′ = (e′ ₁ , e′ ₂ , …, e′ _j , …, e′ _N ), and then the threshold sequence ε is calculated:

ε＝μ(e′)+Z·σ(e′)

计算f(ε_t)：Among them, μ(·) is the mean, σ(·) is the standard deviation, Z represents the artificially set weight, and the threshold sequence ε can be calculated from the reconstruction error sequence e according to the above formula.

Calculate f(ε _t ):

Δμ(e′)=μ(e′)-μ({e∈e′|e<ε _t })

Δσ(e′)=σ({e∈e′|e<ε _t })

e _a ={e∈e′|e＞ε _t }

Wherein, E _seq is a set of continuous data in _ea , the threshold ε that makes f(ε _t ) reach the maximum value is the required threshold, and anomaly detection of time series data is performed based on ε.

Step six, detecting the state of the gas pressure regulator timing data according to the threshold ε, and outputting the normal operation state or abnormal operation state of the gas pressure regulator.

In order to more comprehensively analyze and evaluate the anomaly detection results of the gas pressure regulator time series data, this paper uses precision, recall and F1 value as evaluation indicators. Precision refers to the proportion of data predicted to be abnormal in the data that is actually abnormal; recall refers to the proportion of data predicted to be abnormal in the data that is actually abnormal; F1 value is the harmonic mean of precision and recall. The calculation formulas for the three evaluation indicators are:

Among them, Precision is the precision, Recall is the recall, TP represents the number of abnormal data predicted as abnormal data; FP represents the number of normal data predicted as abnormal data; FN represents the number of abnormal data predicted as normal data.

The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims (10)

1. The method for detecting the abnormality of the medium-low pressure gas pressure regulator based on the LSTM-AE-DT model is characterized by comprising the following steps of:

s1, collecting outlet pressure data of a medium-low pressure gas pressure regulator;

s2, preprocessing the outlet pressure data of the gas pressure regulator to obtain training data and test data;

s3, building an LSTM-AE model, inputting preprocessed training data to train the model, and storing the trained model;

s4, inputting the preprocessed test data into a trained algorithm model, and outputting a predicted value of the outlet pressure of the gas pressure regulator;

s5, reconstructing an error between a predicted value of the output gas pressure regulator outlet pressure and a true value of the gas pressure regulator outlet pressure, and a threshold epsilon of the time sequence data;

s6, detecting the state of the time sequence data of the gas pressure regulator according to the threshold epsilon, and outputting the normal operation state or the abnormal operation state of the gas pressure regulator.

2. The abnormality detection method for the medium-low pressure fuel gas pressure regulator based on the LSTM-AE-DT model of claim 1, wherein in the step S2, the fuel gas pressure regulation is obtained through the step S1Raw outlet pressure time series training data set x= (X) ₁ ,x ₂ ,…,x _j ,…,x _N ) Calculating the average value of the outlet pressure values at the same time point three days before the value of the missing data, and filling the value of the missing data; processing the original time sequence data set obtained in the step S1 by using a deviation normalization method to obtain a normalized time sequence data set X ' = (X ') ' ₁ ,x′ ₂ ,…,x′ _j ,…,x′ _N ),x′ _j ∈[0,1]The formula is as follows:

wherein x is _j For the j-th raw data, x' _j For the j-th normalized time series data, min _1≤j≤N {x _j The minimum value in the training data set, max _1≤j≤N {x _j -represents the maximum value in the training dataset; finally, framing the data set X' by a sliding window method, wherein the sliding window width is l, l is more than or equal to 1 and less than or equal to N, the sliding step length is set to be 1, and the data set Y= (Y) subjected to framing operation at the moment ₁ ,y ₂ ,…,y _N-l+1 ) Wherein y is _i ＝(x′ _i ,x′ _i+1 ,…,x′ _i+l-1 )，1≤i≤N-l+1。

3. The method for detecting the abnormality of the medium-low pressure gas pressure regulator based on the LSTM-AE-DT model according to claim 1, wherein in the step S3, the LSTM network adopts forgetting gate screening information, an input gate retains necessary information and codes, a hidden state is output through an output gate, and the hidden state is transmitted to the next LSTM unit for training, and the data is subjected to characteristic extraction.

4. The method for detecting the abnormality of the medium-low pressure gas pressure regulator based on the LSTM-AE-DT model according to claim 1, wherein the step S3 constructs the LSTM network on an encoder and a decoder of AE, and the encoder acquires the high-dimensional input data sequence in a vector form with a fixed size.

5. The anomaly detection method for the medium-low pressure gas pressure regulator based on the LSTM-AE-DT model of claim 1, wherein in the step S4, the data y after framing is obtained _i ＝(x′ _i ，x′ _i+1 ，，x′ _i+l-1 ) Inputting the LSTM-AE model for training to obtain a reconstructed time sequence data set

6. The anomaly detection method for a medium-low pressure fuel gas pressure regulator based on an LSTM-AE-DT model as set forth in claim 1, wherein in said step S5, a reconstruction error e of a predicted value of the outlet pressure of the fuel gas pressure regulator and a true value of the outlet pressure of the fuel gas pressure regulator is calculated _j ：

Wherein x' _j Represents the j-th outlet pressure data, n represents x' _j The number of times of the reconstruction,

represents x' _j The kth reconstructed outlet pressure data, represented by x' _j Reconstruction error e of (2) _j The calculation method of (1) obtains a reconstruction error sequence: e= (e) ₁ ，e ₂ ，…，e _j ，…，e _N ) Exponentially weighted sliding average is performed on e to obtain an error sequence e '= (e' ₁ ，e′ ₂ ，…，e′ _j ，…，e′ _N ) The threshold sequence ε is then calculated:

ε＝μ(e′)+Z·σ(e′)

wherein μ (·) is the mean, σ (·) is the standard deviation, Z represents the artificially set weight,

calculating f (epsilon) _t )：

Δμ(e′)＝μ(e′)-μ({e∈e′|e＜ε _t })

Δσ(e′)＝σ({e∈e′|e＜ε _t })

e _a ＝{e∈e′|e＞ε _t }

Wherein E is _seq E is _a Such that f (epsilon) _t ) The threshold epsilon at which the maximum value is reached is a threshold value, and abnormality detection of time series data is performed based on epsilon.

7. The method for detecting the abnormality of the medium-low pressure gas pressure regulator based on the LSTM-AE-DT model according to claim 1, wherein in the step S5, the LSTM-AE-DT model is optimized by using a back propagation algorithm, and the network parameters are updated until the input and output loss function L reaches a minimum value, and the loss function L has a formula:

wherein e _j Is the reconstruction error of the predicted value of the outlet pressure of the gas pressure regulator and the true value of the outlet pressure of the gas pressure regulator.

8. The method for detecting the abnormality of the medium-low pressure gas pressure regulator based on the LSTM-AE-DT model according to claim 1, wherein in the step S6, the accuracy, the recall rate and the F1 value are adopted as evaluation indexes, and the calculation formula is as follows:

precision is the Precision, recall is the Recall, TP represents the number of abnormal data predicted as abnormal data; FP represents the number of abnormal data predicted from normal data; FN represents the number of abnormal data predicted as normal data.

9. The method for detecting the abnormality of the medium-low pressure gas pressure regulator based on the LSTM-AE-DT model according to claim 1, wherein the LSTM-AE model in step S3 uses the storage capability of the LSTM to make the data processed by the encoder maintain long dependency of time series data, and simultaneously compress the high-dimensional input vector to the low-dimensional vector.

10. The method for detecting the abnormality of the medium-low pressure gas pressure regulator based on the LSTM-AE-DT model according to claim 1, wherein the step S5 is characterized in that a dynamic threshold method is used for calculating the threshold epsilon of the time series data.

Images