CN110516735A - A Natural Gas Pipeline Event Classification Method Based on LSTM Network and Adam Algorithm - Google Patents

Abstract

Translated fromChinese

本发明涉及一种基于LSTM网络和Adam算法的天然气管道事件分类方法，属于天然气管道在线安全监测领域。该过程包括：首先提取原始样本并进行归一化处理及标签标记，然后分成训练集和验证集。其次，构建LSTM网络模型并将归一化后样本输入网络进行训练及验证。当网络训练精度满足要求时，输出网络模型；否则，重新进行网络训练。最后，将归一化后的新样本输入训练完毕的网络模型，即可实现天然气管道水合物堵塞、泄漏、弯道等事件的分类识别。本发明的优点在于，采用LSTM网络和Adam算法进行管道事件分类无需进行特征提取预处理，且网络训练速度快，能够准确识别管道事件。

The invention relates to a natural gas pipeline event classification method based on an LSTM network and an Adam algorithm, and belongs to the field of online safety monitoring of natural gas pipelines. The process includes: first extracting the original samples and performing normalization processing and labeling, and then divided into training set and verification set. Secondly, construct the LSTM network model and input the normalized samples into the network for training and verification. When the network training accuracy meets the requirements, output the network model; otherwise, retrain the network. Finally, input the normalized new samples into the trained network model to realize the classification and identification of natural gas pipeline hydrate blockage, leakage, bends and other events. The invention has the advantages that the LSTM network and the Adam algorithm are used to classify the pipeline events without feature extraction preprocessing, and the network training speed is fast, and the pipeline events can be accurately identified.

Description

Translated fromChinese

一种基于LSTM网络和Adam算法的天然气管道事件分类方法A Natural Gas Pipeline Event Classification Method Based on LSTM Network and Adam Algorithm

技术领域technical field

本发明涉及一种基于LSTM网络和Adam算法的天然气管道事件分类方法，属于天然气管道在线安全监测领域。The invention relates to a natural gas pipeline event classification method based on an LSTM network and an Adam algorithm, and belongs to the field of online safety monitoring of natural gas pipelines.

背景技术Background technique

由于天然气具有低污染和高热值等优点，在能源消费结构中占比逐年提升。管道作为天然气运输的主要方式之一，其安全运行已受到广泛关注。然而，在高压和低温环境下，天然气水合物极易生成从而影响天然气的正常运输。由于腐蚀、堵塞或第三方施工导致的管道泄漏事故同样会造成严重的安全事故。此外，弯道对于基于声学的管道异常事件探测与定位存在干扰。因此，对天然气管道事件分类识别可为管道运营方采取相应维护措施提供有效指导。Due to the advantages of low pollution and high calorific value, the proportion of natural gas in the energy consumption structure has increased year by year. As one of the main modes of natural gas transportation, the safe operation of pipelines has received extensive attention. However, under high pressure and low temperature environment, natural gas hydrate is very easy to form, which affects the normal transportation of natural gas. Pipeline leakage accidents due to corrosion, blockage or third-party construction can also cause serious safety accidents. In addition, the curve interferes with the acoustic-based detection and location of pipeline anomalies. Therefore, the classification and identification of natural gas pipeline events can provide effective guidance for pipeline operators to take corresponding maintenance measures.

目前，国内外已开展了管道异常事件探测与定位方法的研究并公布了相应成果。一种基于声发射技术被提出来并用于气体水合物流动规律和结晶过程的监测，该技术根据声信号的振幅可以检测到晶体的团聚和形成。一个基于压力波传播法检测天然气管道水合物堵塞系统被设计用于检测堵塞，实现了单水合物堵塞的定位及堵塞程度判定。针对管道泄漏，一种基于观测器和混合整数偏微分方程约束优化的天然气管道多点泄漏检测方法被提出，该离散化方法可大大减少泄漏位置求解时的计算量。一种基于核主成分分析和支持向量机的泄漏检测方法被提出。通过核主成分分析实现声信号的特征提取，然后采用支持向量机实现泄漏等级的识别。然而，上述方法不能实现不同类型管道事件的同时探测与定位。为此，一种基于主动声学激励技术被提出并应用于天然气管道水合物堵塞在线监测。之后，该技术也被证明适用于管道泄漏的监测。此外，基于小波包分析的“能量-模式”方法及混沌特性分析法被提出用于水合物堵塞和管道泄漏的分类识别。为了在保证系统空间分辨率的前提下拓展系统监测范围，基于声学脉冲压缩的天然气管道安全监测方法被提出，该方法较好地解决了监测范围与空间分辨率之间的矛盾。然而，经过匹配滤波后的信号特征不明显，应用传统方法进行特征提取和分类识别效果较差。At present, the research on the detection and location methods of pipeline abnormal events has been carried out at home and abroad, and the corresponding results have been published. A technology based on acoustic emission was proposed and used to monitor gas hydrate flow and crystallization process. This technology can detect the agglomeration and formation of crystals according to the amplitude of the acoustic signal. A natural gas pipeline hydrate blockage detection system based on the pressure wave propagation method is designed to detect blockage, which realizes the location of monohydrate blockage and the judgment of blockage degree. Aiming at pipeline leakage, a multi-point leak detection method of natural gas pipeline based on observer and mixed integer partial differential equation constraint optimization is proposed. This discretization method can greatly reduce the amount of calculation when solving the leak location. A leak detection method based on kernel principal component analysis and support vector machine is proposed. The feature extraction of the acoustic signal is realized by kernel principal component analysis, and then the recognition of the leakage level is realized by using the support vector machine. However, the above methods cannot realize the simultaneous detection and location of different types of pipeline events. For this reason, a technology based on active acoustic excitation was proposed and applied to the online monitoring of natural gas pipeline hydrate blockage. Later, the technology was also proven suitable for monitoring pipeline leaks. In addition, the "energy-mode" method based on wavelet packet analysis and the chaotic characteristic analysis method are proposed for the classification and identification of hydrate blockage and pipeline leakage. In order to expand the monitoring range of the system on the premise of ensuring the spatial resolution of the system, a natural gas pipeline safety monitoring method based on acoustic pulse compression was proposed, which better resolved the contradiction between the monitoring range and the spatial resolution. However, the signal features after matched filtering are not obvious, and the effect of feature extraction and classification recognition using traditional methods is poor.

因此，将LSTM网络应用于天然气管道事件分类，可无需进行特征提取预处理，直接将一维时间序列样本输入网络进行训练。在网络训练过程中，采用Adam算法加速网络收敛。该方法可以实现水合物堵塞及管道泄漏的高精度分类识别，同时准确排除弯道干扰，可为管道运营方采取相应维护措施提供有效指导。Therefore, applying the LSTM network to natural gas pipeline event classification can directly input one-dimensional time series samples into the network for training without the need for feature extraction preprocessing. In the network training process, the Adam algorithm is used to accelerate the network convergence. This method can realize high-precision classification and identification of hydrate blockage and pipeline leakage, and at the same time accurately eliminate bend interference, which can provide effective guidance for pipeline operators to take corresponding maintenance measures.

发明内容Contents of the invention

本发明的目的是提供一种基于LSTM网络和Adam算法的天然气管道事件分类方法，该技术可有效排除弯道干扰，实现水合物堵塞及管道泄漏的分类识别。The purpose of the present invention is to provide a natural gas pipeline event classification method based on LSTM network and Adam algorithm, which can effectively eliminate bend interference and realize the classification and identification of hydrate blockage and pipeline leakage.

本发明的技术方案为：一种基于LSTM网络和Adam算法的天然气管道事件分类方法，包括如下步骤：The technical scheme of the present invention is: a kind of natural gas pipeline event classification method based on LSTM network and Adam algorithm, comprises the following steps:

1)将天然气管道安全监测系统采集的不同事件反射信号经过匹配滤波后作为原始样本，对原始样本进行标签标记，并分成训练集和测试集；1) The reflection signals of different events collected by the natural gas pipeline safety monitoring system are matched and filtered as the original samples, and the original samples are tagged and divided into a training set and a test set;

2)构建LSTM网络，包括：输入层、LSTM层、全连接层及输出层。设定网络的各项参数，包括：LSTM层数、LSTM隐含层神经元数、全连接层层数、全连接层激活函数、迭代次数；2) Construct the LSTM network, including: input layer, LSTM layer, fully connected layer and output layer. Set various parameters of the network, including: LSTM layer number, LSTM hidden layer neuron number, fully connected layer layer number, fully connected layer activation function, iteration number;

3)将原始样本归一化后输入构建完毕的网络进行训练与验证。网络训练过程中，单个LSTM细胞运行过程如下：3) Normalize the original samples and input them into the constructed network for training and verification. During network training, a single LSTM cell operates as follows:

(1)将当前时间步t的输入值x^<t>与上一时间步t-1的激活值a^<t-1>组合，可表示为：(1) Combine the input value x^<t> of the current time step t with the activation value a^<t-1> of the previous time step t-1, which can be expressed as:

(2)在时间步t的候选细胞状态为：(2) The candidate cell state at time step t is:

(3)求取更新门i^<t>、遗忘门f^<t>及输出门o^<t>：(3) Calculate update gate i^<t> , forget gate f^<t> and output gate o^<t> :

i^<t>＝σ(W_ux_con+b_u) (3)i^<t> ＝σ(W_u x_con +b_u ) (3)

f^<t>＝σ(W_fx_con+b_f) (4)f^<t> = σ(W_f x_con +b_f ) (4)

o^<t>＝σ(W_ox_con+b_o) (5)o^<t> ＝σ(W_o x_con +b_o ) (5)

其中，W为权值矩阵，b为偏置，σ代表更新门、遗忘门和输出门的激活函数。Among them, W is the weight matrix, b is the bias, and σ represents the activation function of the update gate, forget gate and output gate.

(4)遗忘门f^<t>和更新门i^<t>可被用于确定在时间步t-1的细胞状态C^<t-1>及候选细胞状态是否被保留。因此，当前细胞状态C^<t>可被更新为：(4) Forget gate f^<t> and update gate i^<t> can be used to determine cell state C^<t-1> and candidate cell states at time step t-1 whether to be retained. Therefore, the current cell state C^<t> can be updated as:

其中，符号“×”代表Hadamard乘法。Among them, the symbol "×" represents Hadamard multiplication.

(5)更新在时间步t的激活值：(5) Update the activation value at time step t:

a^<t>＝o^<t>×tanhC^<t> (7)a^<t> = o^<t> ×tanhC^<t> (7)

4)将多个LSTM细胞按时间步串联成一个LSTM层后，通过全连接层转换LSTM层输出向量的维度并转换为概率分布，则输出层的概率分布可表示为：4) After concatenating multiple LSTM cells into an LSTM layer according to the time step, the dimension of the output vector of the LSTM layer is converted through the fully connected layer and converted into a probability distribution, then the probability distribution of the output layer can be expressed as:

其中，a为全连接层的输出向量，a_j为该向量中的第j个元素，n为标签状态数。Among them, a is the output vector of the fully connected layer, a_j is the jth element in the vector, and n is the number of label states.

为了训练网络，还需引入交叉熵作为单个样本的损失函数，可表示为：In order to train the network, it is also necessary to introduce cross entropy as the loss function of a single sample, which can be expressed as:

其中，y为样本的真实类别，代表一个标签编码向量。对于N个样本，总交叉熵损失Loss可表示为：Among them, y is the true category of the sample, representing a label encoding vector. For N samples, the total cross-entropy loss Loss can be expressed as:

LSTM网络训练的目的是利用指定的优化算法，通过不断迭代来调整权值W和偏置b，使得交叉熵损失函数最小。The purpose of LSTM network training is to use the specified optimization algorithm to adjust the weight W and bias b through continuous iteration to minimize the cross-entropy loss function.

5)采用Adam算法调整网络权值W和偏置b，使得交叉熵损失函数最小。Adam算法执行过程如下：5) The Adam algorithm is used to adjust the network weight W and the bias b to minimize the cross-entropy loss function. The implementation process of the Adam algorithm is as follows:

(1)求解梯度的一阶矩估计V_dW、V_db和二阶矩估计S_dW、S_db：(1) Solve the first-order moment estimation V_dW , V_db and the second-order moment estimation S_dW , S_db of the gradient:

V_dW：＝β₁V_dW+(1-β₁)dW (11)V_dW : = β₁ V_dW + (1-β₁ )dW (11)

V_db：＝β₁V_db+(1-β₁)db (12)V_db :=β₁ V_db +(1−β₁ )db (12)

S_dW：＝β₂S_dW+(1-β₂)dW² (13)S_dW : = β₂ S_dW + (1-β₂ )dW² (13)

S_db：＝β₂S_db+(1-β₂)db² (14)S_db :=β₂ S_db +(1−β₂ )db² (14)

其中，β₁和β₂一阶、二阶矩估计的指数衰减率，右侧的矩阵V和S为上一次迭代的矩估计，符号“：＝”代表赋值操作。Among them, β₁ and β₂ are the exponential decay rates of the first-order and second-order moment estimates, the matrices V and S on the right are the moment estimates of the last iteration, and the symbol ":=" represents the assignment operation.

(2)修正一阶矩估计和二阶矩估计：(2) Modified first-order moment estimation and second-order moment estimation:

其中，k为迭代次数，和分别为β₁和β₂的k次方。Among them, k is the number of iterations, and are the kth powers of β₁ and β₂ , respectively.

(3)通过修正的矩估计更新参数W和b，则更新的权值W和偏置b可表示为：(3) Update the parameters W and b through the modified moment estimation, then the updated weight W and bias b can be expressed as:

其中，α为学习率，默认为0.0001，ε为10^-8。Among them, α is the learning rate, which is 0.0001 by default, and ε is 10^-8 .

6)判断网络精度是否符合要求。若符合，则输出训练完毕的网络模型；否则，重新进行网络训练；6) Judging whether the network accuracy meets the requirements. If so, output the trained network model; otherwise, re-train the network;

7)将带有标签新样本输入精度符合要求的网络模型进行测试，通过网络输出类别与样本标签进行对比来测试网络模型。7) Test the network model with a new sample input accuracy that meets the requirements, and test the network model by comparing the network output category with the sample label.

所述原始样本提取时仅提取匹配滤波后反射信号主瓣，原始样本大小为1×408。When the original sample is extracted, only the main lobe of the reflected signal after the matched filter is extracted, and the size of the original sample is 1×408.

所述归一化方法为Z-score标准化。The normalization method is Z-score standardization.

所述标签标记为One-Hot编码标记。The label marks are One-Hot coded marks.

所述LSTM层数、LSTM隐含层神经元数、全连接层层数和迭代次数均可调，优选LSTM层数为3层，LSTM隐含层神经元数为64个，全连接层层数为2层，迭代次数为30次。The number of layers of the LSTM, the number of neurons in the hidden layer of the LSTM, the number of layers in the fully connected layer and the number of iterations are all adjustable. The number of layers in the preferred LSTM is 3, the number of neurons in the hidden layer of the LSTM is 64, and the number of layers in the fully connected layer is 2 layers, and the number of iterations is 30.

所述更新门、遗忘门和输出门的激活函数Sigmoid函数。The activation function Sigmoid function of the update gate, the forget gate and the output gate.

所述中间全连接层的激活函数可调，优选Relu函数，顶层全连接层的激活函数为Softmax函数。The activation function of the middle fully connected layer is adjustable, preferably the Relu function, and the activation function of the top fully connected layer is a Softmax function.

本发明的第一个优点在于利用LSTM网络可无需进行样本特征提取预处理即可实现天然气管道事件高精度分类识别；第二个优点在于利用Adam算法可加快网络训练速度。The first advantage of the present invention is that the LSTM network can be used to realize high-precision classification and identification of natural gas pipeline events without sample feature extraction and preprocessing; the second advantage is that the use of Adam algorithm can speed up network training.

附图说明Description of drawings

图1为本发明的过程流程图。Fig. 1 is the process flowchart of the present invention.

图2为本发明的LSTM网络结构图。FIG. 2 is a structural diagram of the LSTM network of the present invention.

图3为本发明的管道事件样本图。Fig. 3 is a sample diagram of a pipeline event in the present invention.

图4为本发明的训练结果图。Fig. 4 is a graph of training results of the present invention.

图5为本发明的测试结果图。Fig. 5 is a diagram of test results of the present invention.

具体实施方式Detailed ways

下面结合附图和具体实施方式对本发明作进一步详细的说明：Below in conjunction with accompanying drawing and specific embodiment the present invention will be described in further detail:

图1所示为本发明的过程流程图。具体过程如下：A、提取原始样本；B、对原始样本进行归一化处理及标签标记；C、构建LSTM网络并设置网络参数；D、进行网络训练与验证，其中全部样本的80％作为训练集，20％作为验证集；E、判断网络精度是否符合要求；若否，则返回步骤C；若是，则进行步骤F；F、输出精度符合要求的网络模型；G、利用归一化后的新样本进行网络测试。Fig. 1 shows the process flow chart of the present invention. The specific process is as follows: A. Extract original samples; B. Normalize and label the original samples; C. Build LSTM network and set network parameters; D. Perform network training and verification, and 80% of all samples are used as training Set, 20% as the verification set; E, judge whether the network accuracy meets the requirements; if not, return to step C; if so, proceed to step F; F, output the network model whose accuracy meets the requirements; G, use the normalized New samples for network testing.

图2所示为LSTM网络结构，包络一个输入层、三个LSTM层、两个全连接层及输出层。归一化后样本可直接输入LSTM网络输入层。然后，经过LSTM层，可捕获输入样本序列的长时间相关关系。在LSTM层后接一个中间全连接层，其激活函数为Relu函数，用于转换LSTM层的输出维度。最后，在顶层再接一个全连接层，其激活函数为Softmax函数，用于转换中间全连接层的输出至管道事件的概率分布。输出层则将顶层全连接层输出的概率分布直接输出。Figure 2 shows the LSTM network structure, enveloping an input layer, three LSTM layers, two fully connected layers and an output layer. The normalized samples can be directly input into the input layer of the LSTM network. Then, through the LSTM layer, the long-term correlation relationship of the input sample sequence can be captured. An intermediate fully connected layer is connected after the LSTM layer, and its activation function is the Relu function, which is used to convert the output dimension of the LSTM layer. Finally, a fully connected layer is connected to the top layer, and its activation function is the Softmax function, which is used to convert the output of the intermediate fully connected layer to the probability distribution of pipeline events. The output layer directly outputs the probability distribution output by the top fully connected layer.

图3(a)-(c)分别为水合物堵塞、管道泄漏和弯道原始样本信号。这三类事件除幅值不同外，波形几乎一致。为了降低管道事件分类识别难度，首先对原始样本数据进行归一化预处理。Z-Score归一化结果如图3(d)-(e)所示，三类样本的波形和幅值均存在明显差异，有效降低了管道事件分类识别的难度。Figure 3(a)-(c) are the original sample signals of hydrate blockage, pipeline leakage and bend, respectively. The waveforms of these three types of events are almost the same except for the different amplitudes. In order to reduce the difficulty of classification and identification of pipeline events, the original sample data is firstly normalized and preprocessed. The Z-Score normalization results are shown in Figure 3(d)-(e). There are obvious differences in the waveforms and amplitudes of the three types of samples, which effectively reduces the difficulty of classification and identification of pipeline events.

图4所示为训练过程中交叉熵损失和训练精度。图4(a)和(b)的横轴均代表迭代次数，纵轴分别代表交叉熵损失与训练精度。在网络训练过程中，交叉熵损失函数快速且稳定收敛。同时，训练精度也稳步提升。当达到指定迭代次数后，交叉熵损失为0.0005，训练精度为100.00％。为了验证网络训练精度以调整参数，将验证集输入训练模型，训练精度为100％。因此，该模型精度符合要求。Figure 4 shows the cross-entropy loss and training accuracy during training. The horizontal axes of Figure 4(a) and (b) both represent the number of iterations, and the vertical axes represent the cross-entropy loss and training accuracy, respectively. During network training, the cross-entropy loss function converges quickly and stably. At the same time, the training accuracy is also steadily improved. When the specified number of iterations is reached, the cross-entropy loss is 0.0005, and the training accuracy is 100.00%. In order to verify the network training accuracy to adjust the parameters, the validation set is fed into the training model, and the training accuracy is 100%. Therefore, the accuracy of the model meets the requirements.

图5所示为网络测试图。将未参与训练和验证且归一化后的30个新样本输入训练完毕且精度符合要求的网络模型进行测试。测试结果如图5所示，柱状线代表输出概率分布，与新样本标记的概率分布接近。因此，可将这些新样本正确地分类为水合物堵塞、管道泄漏和弯道，分类精度为100％，证明了该网络模型可被有效应用于天然气管道事件的分类。Figure 5 shows the network test diagram. The normalized 30 new samples that have not participated in training and verification are input into the network model that has been trained and the accuracy meets the requirements for testing. The test results are shown in Figure 5, and the column line represents the output probability distribution, which is close to the probability distribution of the new sample label. Therefore, these new samples can be correctly classified as hydrate blockage, pipeline leakage and bend with 100% classification accuracy, proving that the network model can be effectively applied to the classification of natural gas pipeline events.

Claims (7)

1. a kind of natural gas line event category method based on LSTM network and Adam algorithm, it is characterised in that including following mistakeJourney:

1) the different event reflection signal acquired natural gas line safety monitoring system is after matched filtering as original sampleThis, carries out label label to original sample, and be divided into training set and test set；

2) LSTM network is constructed, comprising: input layer, LSTM layers, full articulamentum and output layer.The parameters of setting network, packetIt includes: the LSTM number of plies, LSTM hidden layer neuron number, the full articulamentum number of plies, full articulamentum activation primitive, the number of iterations；

3) network that input building finishes after normalizing original sample is trained and verifies.In network training process, individuallyLSTM cell operational process is as follows:

(1) by the input value x of current time step t^<t>With the activation value a of upper time step t-1^<t-1>Combination, may be expressed as:

(2) in the candidate cell state of time step t are as follows:

(3) it seeks updating door i^<t>, forget door f^<t>And out gate o^<t>:

i^<t>=σ (W_ux_con+b_u) (3)

f^<t>=σ (W_fx_con+b_f) (4)

o^<t>=σ (W_ox_con+b_o) (5)

Wherein, W is weight matrix, and b is biasing, and σ represents the activation primitive for updating door, forgeing door and out gate.

(4) forget door f^<t>With update door i^<t>It can be used to determine the cell state C in time step t-1^<t-1>And candidate cell shapeStateWhether it is retained.Therefore, current cell state C^<t>It can be updated to:

Wherein, symbol "×" representsHadamard multiplication.

(5) activation value in time step t is updated:

a^<t>=o^<t>×tanhC^<t> (7)

4) by multiple LSTM cells after spacer step is connected into one LSTM layers on time, pass through full articulamentum and convert LSTM layers of output vectorDimension and be converted to probability distribution, then the probability distribution of output layer may be expressed as:

Wherein, a is the output vector of full articulamentum, a_jFor j-th of element in the vector, n is tag state number.

In order to train network, loss function of the cross entropy as single sample is introduced, may be expressed as:

Wherein, y is the true classification of sample, represents a label coding vector.It is total to intersect entropy loss Loss for N number of sampleIt may be expressed as:

The purpose of LSTM network training is, using specified optimization algorithm, weight W and biasing b to be adjusted by continuous iteration, is madeIt is minimum to obtain cross entropy loss function.

5) using Adam algorithm adjustment network weight W and biasing b, so that cross entropy loss function is minimum.Adam algorithm executedJourney is as follows:

(1) the single order moments estimation V of gradient is solved_dW、V_dbWith second order moments estimation S_dW、S_db:

V_dW:=β₁V_dW+(1-β₁)dW (11)

V_ab:=β₁V_db+(1-β₁)db (12)

S_dW:=β₂S_dW+(1-β₂)dW² (13)

S_ab:=β₂S_db+(1-β₂)db² (14)

Wherein, β₁And β₂The exponential decay rate of single order, second order moments estimation, the matrix V and S on right side are the moments estimation of last iteration,Symbol " :=" represent assignment operation.

(2) single order moments estimation and second order moments estimation are corrected；

Wherein, k is the number of iterations,WithRespectively β₁And β₂K power.

(3) by modified moments estimation undated parameter W and b, then the weight W updated and biasing b may be expressed as:

Wherein, α is learning rate, is defaulted as 0.0001, ε 10^-8。

6) judge whether neural network accuracy meets the requirements.If meeting, the network model that training finishes is exported；Otherwise, it re-startsNetwork training；

7) it will be tested with the satisfactory network model of label new samples input precision, classification and sample are exported by networkThis label, which compares, carrys out test network model.

2. a kind of natural gas line event category method based on LSTM network and Adam algorithm as described in claim 1,Be characterized in that: the original sample only extracts matched filtering back reflection signal main lobe when extracting, and original sample size is 1 × 408.

3. a kind of natural gas line event category method based on LSTM network and Adam algorithm as described in claim 1,Be characterized in that: the method for normalizing is Z-score standardization.

4. a kind of natural gas line event category method based on LSTM network and Adam algorithm as described in claim 1,Be characterized in that: the label is labeled as One-Hot coded markings.

5. a kind of natural gas line event category method based on LSTM network and Adam algorithm as described in claim 1,Be characterized in that: the LSTM number of plies, LSTM hidden layer neuron number, the full articulamentum number of plies and the number of iterations are adjustable, preferablyThe LSTM number of plies is 3 layers, and LSTM hidden layer neuron number is 64, and the full articulamentum number of plies is 2 layers, and the number of iterations is 30 times.

6. a kind of natural gas line event category method based on LSTM network and Adam algorithm as described in claim 1,Be characterized in that: the activation primitive for updating door, forgeing door and out gate is Sigmoid function.

7. a kind of natural gas line event category method based on LSTM network and Adam algorithm as described in claim 1,Be characterized in that: the activation primitive of the full articulamentum in centre is adjustable, preferably Relu function, and the activation primitive of the full articulamentum of top layer isSoftmax function.