Nuclear power circulating water pump on-line abnormal monitoring and identification method

Abstract

A nuclear power circulating water pump online abnormality monitoring and identification method is disclosed. In the method, multi-source time series data in the circulating water pump operation process is obtained online, data labels are given through an expert analysis module, data normalization processing is performed, and signal is randomly divided into training set and test set samples. When the input data stream only contains healthy data, a data fusion monitoring module based on a feature layer constraint long short-time self-encoder, an attention mechanism and a gate structure can fully mine the state information and potential correlation characteristics of the data, and realize state monitoring under healthy samples. With the occurrence of abnormal data in the monitoring data, an abnormality identification network is further constructed based on the encoder feature, a joint loss function based on cross-entropy, distillation loss and triple loss is constructed, and a neighboring mean value classifier is used to realize online monitoring and identification of known / unknown abnormal states. The method has good compatibility and expansibility, and pushes more accurate state monitoring results to operation and maintenance personnel.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of online monitoring of nuclear power circulating water pumps, and particularly relates to a method for online abnormal monitoring and identification of nuclear power circulating water pumps. BACKGROUND

[0002] At present, the main means for nuclear power plants to evaluate the state of circulating water pumps is still the threshold alarm-based method, which has serious defects such as missed alarms, false alarms, and reaction lag. With the increasing maturity of artificial intelligence technology, online monitoring models based on deep learning can identify existing abnormalities in equipment and improve the safety of operation. In recent years, many scholars at home and abroad have applied deep learning to evaluate the operating state of key components of equipment and have achieved good results. However, unsupervised state monitoring methods cannot utilize possible abnormal samples, and can only judge health or abnormality, so maintenance personnel still need to carry out a lot of maintenance to locate the fault source.

[0003] In order to realize the state monitoring of the planetary gear box of the circulating water pump and further identify the known abnormality, the existing method mainly relies on a complete offline data set to train the monitoring model and the classification model at the same time. As the service life of the circulating water pump increases, new abnormal data will be generated, which will cause the original model to fail, thereby increasing the application cost of the method and reducing the flexibility. Moreover, the single-source data of the existing technology has great limitations for the overall state evaluation of the planetary gear box of the nuclear power circulating water pump and its associated components.

[0004] The above information disclosed in the background section is only intended to enhance the understanding of the background of the present application, and therefore can contain information that is not prior art known to those of ordinary skill in the art. SUMMARY

[0005] The purpose of the present application is to provide a method for online abnormal monitoring and identification of nuclear power circulating water pumps. When the input data stream only contains healthy data, the present application designs a data fusion monitoring module based on a self-encoder of a long short-term memory network. The attention mechanism and the gate structure can fully exploit the state information and potential correlation characteristics of the data, and realize the state monitoring under healthy samples. With the emergence of abnormal data in the monitoring data, an abnormal identification network is further constructed based on the encoder features, a joint loss function based on cross-entropy, distillation loss and triplet loss is constructed to ensure that the model retains the memory of the old data when learning new data. Finally, a nearest mean classifier is used to realize online monitoring and identification of known / unknown abnormal states. The proposed method has good compatibility and expandability, can be applied to various states of equipment operation, and can further remember and distinguish known and unknown abnormalities, and push more accurate state monitoring results to the maintenance personnel.

[0006] In order to achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0007] The nuclear power plant off-site emergency evacuation path planning method comprises the following steps:

[0008] The first step is to obtain multi-source time series data in the operation process of the nuclear power circulating water pump online and give data labels, and then the multi-source time series data is sequentially sampled with a predetermined length without overlapping, and then divided into training set samples or test set samples;

[0009] The second step is to collect healthy samples before the operation of the nuclear power circulating water pump, and to construct a long-short time autoencoder to realize state monitoring under the healthy samples, wherein the maximum mean difference MMD is used as a regularization term to align the encoding layer feature distribution and the Gaussian distribution, and the given latent representation is selected in the reproducing kernel Hilbert space H z The maximum mean difference MMD is calculated as follows:

[0010]

[0011] wherein, is the feature of the sample z i ; z (i) is the sampling result of the target distribution as the Gaussian distribution; φ represents the mapping from the sample feature space to the reproducing kernel Hilbert space H, and the low-dimensional representation is the distribution Z and the target distribution z (i) ~ P z on the set X, and the maximum mean difference MMD is defined as the feature mapping φ: X→H, wherein H is a reproducing kernel Hilbert space, and the maximum mean difference MMD on the set X is calculated in one iteration, wherein the number of X is equal to the product of the training batch size B and the data length;

[0012] The third step is to monitor the occurrence of abnormal data in the operation process of the nuclear power circulating water pump, to construct an anomaly recognition network based on the encoder feature, and to construct a joint loss function based on cross-entropy, distillation loss and triplet loss to make the memory of the old data be retained when learning new data, wherein the encoder of the long-short time autoencoder is constructed as a feature extractor of the anomaly recognition network, and in the joint loss function, the cross-entropy L CE is:

[0013]

[0014] wherein, x i is the sample input; y is the sample category; y i is the predicted label of the i-th sample; s is the sample label of the new category; log is an exponential function; g y is a classifier; and δ is a discriminant function, and y iOutput is 1, otherwise 0;

[0015] Distillation loss L DL For reproducing the output of old samples in the new model in the original model to suppress the forgetting of learned knowledge:

[0016]

[0017] Where q i is the predicted output of the old sample by the original model; s-1 represents the number of learned classes;

[0018] Triplet loss function Loss TN is:

[0019] L TN = max(0, ||N(x1)-N(x2)|| 2 -||N(x1)-N(x3)|| 2 +m),

[0020] Where N(x i ) is the output of the neural network for sample x i ; m is the minimum boundary distance set, x1 of class C is called anchor, x2 from class C is called positive sample, and x3 from other classes is called negative sample respectively;

[0021] The fourth step is to use the adjacent mean classifier to online monitor and identify the real-time data of the nuclear power circulating water pump. The abnormal identification network outputs the yth class of health state sample features, and the average feature vector μ y is calculated:

[0022]

[0023] Where P y is the example set of class y; is the mapping function of the feature extractor; m is the number of samples of class y;

[0024] The Euclidean distance is used to distinguish the sample classes, and the sample to be classified is classified into the class closest to the average feature vector:

[0025]

[0026] Where t is the total number of final learned classes; is the class prediction value;

[0027] The reconstruction error based on the health sample is used to set the abnormal monitoring threshold. When the reconstruction error is higher than the abnormal monitoring threshold, it is determined as abnormal data.

[0028] The multi-source time sequence data includes motor base Z-direction vibration acceleration, gear box body Z-direction vibration acceleration signal, pump guide bearing Z-direction vibration acceleration, gear box output shaft nearby sound pressure, gear box output shaft X-direction displacement and gear box output shaft Y-direction displacement.

[0029] In the nuclear power circulating water pump online abnormality monitoring and identification method, the predetermined length is 2048, the length of the segmented single data is 2048, the channel is 6 channels, and the single data is 6*2048 two-dimensional data.

[0030] In the nuclear power circulating water pump online abnormality monitoring and identification method, for the samples x1, x2, x3, …, x n The following transformation is performed:

[0031]

[0032] Among them, is the minimum data value of the corresponding sample, is the maximum data value in the corresponding sample, and each sample is normalized in the interval [0, 1].

[0033] In the nuclear power circulating water pump online abnormality monitoring and identification method, the long short-term memory neural network includes a gate unit and a storage unit, the long short-term memory neural network generates a vector and selects candidate information between 0 and 1, in the input gate, the vector I t is generated by the hidden state h t-1 and the input x t of the current step, and f g , f i , f c , f o , f h are different stage activation functions, w i , w g , w o are weight matrices, and b i , b g , b o are biases, and the generation process can be represented as:

[0034] I t =f i (w i x t +w i hl t-1 +b i ),

[0035] The forget gate F t determines the state C t-1Whether it should be retained, expressed as:

[0036] F t = f g (w g x t +w g h t-1 +b g ),

[0037] Then, the input candidate information is generated by x t and h t-1

[0038]

[0039] Further, the current cell state C t is obtained as:

[0040]

[0041] The output gate also generates a vector Y t :

[0042] Y t = f o (w o x t +w o h t-1 +b o ),

[0043] Finally, the hidden state h t is obtained according to Y t and C t :

[0044] h t =Y t f h (C t ),

[0045] The mean square error is used as the loss function to realize the training of the long short-term memory neural network, and the mean square error L MSE is expressed as:

[0046]

[0047] Wherein, N is the total number of samples of all categories; x t is the model input data; is the reconstructed data.

[0048] In the nuclear power circulating water pump online abnormal monitoring and identification method, in the training process, the maximum mean difference MMD is calculated by using the nuclear skill, and k(x, y)=(φ(x), φ(y))​H The maximum mean discrepancy MMD is expressed as:

[0049]

[0050] Gaussian distribution P z The potential representation is approximated by sampling from P z MMD on the set X is computed in one iteration, where the number of X is equal to the product of the training batch size B and the data length.

[0051] The abnormality recognition network comprises a batch normalization layer, and the activation function is set to ReLU.

[0052] The output result update interval is set to 5s.

[0053] The latent space d < N x T.

[0054] The abnormality recognition network is trained by using a back propagation algorithm, the gradient of each weight is calculated by using chain rule, the weight is updated according to the gradient descent algorithm, and the cost function used by the convolutional neural network is a cross entropy function, and the formula is:

[0055]

[0056] Wherein, c represents the cost, x represents the sample, n represents the total number of samples, and a represents the model output value.

[0057] In the technical solution, the method for online abnormal monitoring and identification of the nuclear power circulating water pump has the following beneficial effects: the method can obtain the equipment operation data online and further learn the new data, and can fully utilize the healthy samples and the new abnormal samples. The online abnormal state monitoring method uses the long-short self-encoder structure to learn the compression reconstruction of the healthy data unsupervisedly, and determines the abnormal data based on the reconstruction error. In order to suppress the generalization learning of the feature layer feature distribution of the traditional long-short self-encoder and constrain the feature layer distribution, the maximum mean difference is used to constrain the output distribution of the feature layer, so as to reduce the deviation of the feature layer output distribution and the standard Gaussian distribution. The results prove that the method is stable in training and can effectively monitor various types of abnormalities. For the abnormal identification problem, the method simulates the training form of the "data flow" of the industrial field, proposes a lifelong learning architecture to fully utilize the existing healthy samples and abnormal samples generated at any time, and realizes the full utilization of data and online learning and application of the model. The distillation loss and the ternary loss are used as loss function compensation terms to alleviate the forgetting of the model to the old categories and regulate the sample distribution; the use of the example space alleviates the storage pressure of the data acquisition terminal. The test proves that the method performs better than the widely used lifelong learning method icaRL and the fine-tuning method, and verifies its abnormal identification ability under different experimental settings. The method is more suitable for the actual application scene of the industrial field, has good compatibility and expansibility, can be applied to various states of equipment operation, and can further remember and distinguish known and unknown abnormalities, and push more accurate state monitoring results to the operation and maintenance personnel. BRIEF DESCRIPTION OF DRAWINGS

[0058] In order to more clearly illustrate the technical solutions in the embodiments or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0059] Figure 1 The flowchart of the method for online abnormal monitoring and identification of the nuclear power circulating water pump in the present application;

[0060] Figure 2 The structure diagram of the nuclear power circulating water pump scaling experiment in the method for online abnormal monitoring and identification of the nuclear power circulating water pump in the present application;

[0061] Figure 3 The time-frequency diagram of the multi-source data in the method for online abnormal monitoring and identification of the nuclear power circulating water pump in the present application;

[0062] Figure 4 Loss function curve comparison schematic diagram of the online abnormal monitoring and identification method of the circulating water pump in the nuclear power plant in the application;

[0063] Figure 5 Self-encoder reconstruction error distribution diagram of the online abnormal monitoring and identification method of the circulating water pump in the nuclear power plant in the application;

[0064] Figure 6 Long-short self-encoder reconstruction error distribution diagram of the online abnormal monitoring and identification method of the circulating water pump in the nuclear power plant in the application;

[0065] Figure 7 Monitoring result diagram of the online abnormal monitoring and identification method of the circulating water pump in the nuclear power plant in the application;

[0066] Figure 8 Comparison method confusion matrix diagram of the online abnormal monitoring and identification method of the circulating water pump in the nuclear power plant in the application;

[0067] Figure 9 Confusion matrix diagram of the online abnormal monitoring and identification method of the circulating water pump in the nuclear power plant in the application. DETAILED DESCRIPTION

[0068] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in a clear and complete manner with reference to the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0069] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0070] It should be noted that: similar reference numerals and letters represent similar items in the following drawings, so once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0071] In the description of the present application, it needs to be understood that the terms "center", "longitudinal", "transverse", "length", "width", "thickness", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, which is only for the convenience of describing the present application and simplifying the description, and does not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0072] In addition, the terms "first", "second" are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise explicitly specified and limited.

[0073] In the present application, unless otherwise explicitly specified and limited, the terms "mounting", "connecting", "connecting", "fixing" and the like should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or it can be integrated; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be the internal communication of two elements or the interaction relationship between two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0074] In the present application, unless otherwise explicitly specified and limited, the first feature "on" or "under" the second feature can include that the first and second features are in direct contact, or that the first and second features are not in direct contact but are in contact through another feature between them. Moreover, the first feature "on", "above" and "above" the second feature includes that the first feature is directly above and obliquely above the second feature, or only indicates that the horizontal height of the first feature is higher than that of the second feature. The first feature "under", "below" and "below" the second feature includes that the first feature is directly below and obliquely below the second feature, or only indicates that the horizontal height of the first feature is less than that of the second feature.

[0075] In order for those skilled in the art to better understand the technical solutions of the present application, the present application will be further described in detail below. As shown in the drawings, Figures 1 to 9 A nuclear power circulating water pump online abnormality monitoring and identification method includes,

[0076] In a first step, multi-source time series data in the operation of a nuclear power circulating water pump are acquired online and given data labels, the multi-source time series data are sequentially sampled with a predetermined length of non-overlapping sampling, and then divided into training set samples or test set samples;

[0077] In a second step, healthy samples are collected before the operation of the nuclear power circulating water pump, and a long-short term autoencoder is constructed to realize state monitoring under the healthy samples, wherein a maximum mean difference MMD is used as a regularization term to align the feature distribution of the encoding layer and a Gaussian distribution, and a latent representation is given where d is a latent space, h represents all time steps in one iteration, and a Gaussian distribution P is selected in a reproducing kernel Hilbert space H z As a target distribution, the maximum mean difference is calculated as follows:

[0078]

[0079] where, is the feature of a sample x i ; z (i) is a sampling result with a Gaussian distribution as a target distribution; and φ represents a mapping from a sample feature space to a reproducing kernel Hilbert space H

[0080] In a third step, when abnormal data are monitored in the operation of the nuclear power circulating water pump, an anomaly recognition network is constructed based on the encoder features, a joint loss function based on cross-entropy, distillation loss and triplet loss is constructed to enable the memory of old data to be retained when new data are learned, in an example, the sample length is set to 2048, and the sliding window offset is set to 256. The total amount of storable samples is set to K=1000, the number of data reconstruction training rounds is set to 40, the model is evaluated for 60 times in each incremental evaluation, the network training adopts an Adam optimizer, the learning rate is set to 0.001, and the weight decay coefficient is set to 0.00001. In the joint loss function,

[0081] The cross-entropy L CE is:

[0082]

[0083] where x i is a sample input; y is a sample category; y i is a predicted label of the jth sample; s is a sample label of a new category; log is an exponential function; g y is a classifier; and δ is a discriminant function, which outputs 1 only when y=y i , and 0 otherwise.

[0084] The distillation loss L DLReproduce the output of the old sample in the new model in the original model to suppress the forgetting of learned knowledge:

[0085]

[0086] Wherein, q i is the prediction output of the old sample by the original model; s-1 represents the number of learned classes;

[0087] The triplet loss function Loss TN is:

[0088] L TN = max(0, ||N(x1)-N(x2)|| 2 -||N(x1)-N(x3)|| 2 +m),

[0089] Wherein, N(x i ) is the output of the neural network to the sample x i ; m is the minimum boundary distance set, x1 of C class is called anchor, x2 from C class is called positive sample, x3 from other classes is called negative sample respectively;

[0090] The gate unit and storage unit in the structure of Long Short-term Memory Networks (LSTM) can effectively maintain and transmit the key features of data. LSTM-AE combines LSTM network with auto-encoder. This structure combines the powerful time series data feature mining ability of LSTM and the data reconstruction ability of auto-encoder, and can mine the spatio-temporal correlation features of data under the unsupervised setting, realizing the reconstruction of multi-source time series data. LSTM-AE realizes the compression and reconstruction of input signal x t by using the composition of encoder E and decoder D. Further, the mean square error is used as the loss function to realize the training of the model, and the mean square error L MSE can be expressed as:

[0091]

[0092] Wherein, x t is the actual value of the tth sample, is the predicted value of the tth sample;

[0093] To avoid feedback lag caused by low-frequency data, multi-source high-frequency data is used as input, and the selection of measuring points includes pressure generated under the gearbox, vibration acceleration of the motor base, vibration acceleration of the gearbox body, X-direction displacement of the gearbox output shaft, Y-direction displacement of the gearbox output shaft, and vibration acceleration on the upper side of the impeller, totaling 6 time series signals. The signals can reflect the running state of the planetary gearbox of the circulating water pump and related components such as the upper motor, the gearbox output shaft, the gearbox oil pump, and the upper impeller.

[0094] In the fourth step, the adjacent mean value classifier is used for online monitoring and identification of real-time data of the nuclear circulating water pump. The abnormal identification network outputs the y-th health state sample feature, and the average feature vector μ y :

[0095]

[0096] wherein P y is a set of examples of the y-th class; is a mapping function of the feature extractor; and m is the number of samples of the y-th class.

[0097] The Euclidean distance is used to distinguish the sample classes, and the sample to be classified is classified into the class that is most adjacent to the average feature vector:

[0098]

[0099] wherein t is the total number of classes learned finally; is a class prediction value.

[0100] The reconstruction error based on the health samples is used to set an abnormal monitoring threshold. When the reconstruction error is higher than the abnormal monitoring threshold, the data is determined to be abnormal.

[0101] In the preferred embodiment of the online abnormal monitoring and identification method of the nuclear circulating water pump, the multi-source time series data includes Z-direction vibration acceleration of the motor base, Z-direction vibration acceleration of the gearbox body, Z-direction vibration acceleration of the pump guide bearing, sound pressure near the gearbox output shaft, X-direction displacement of the gearbox output shaft, and Y-direction displacement of the gearbox output shaft.

[0102] In the preferred embodiment of the online abnormal monitoring and identification method of the nuclear circulating water pump, the predetermined length is 2048, the length of the segmented single data is 2048, the channel is 6 channels, and the single data is 6x2048 two-dimensional data.

[0103] In the preferred embodiment of the online abnormal monitoring and identification method of the nuclear circulating water pump, the samples x1, x2, x3, …, x n are transformed as follows:

[0104]

[0105] wherein, is the minimum data value of the corresponding sample, is the maximum data value in the corresponding sample, and each sample is normalized in the interval [0, 1].

[0106] In the preferred embodiment of the online abnormal monitoring and identification method of the nuclear power circulating water pump, the long short-term memory neural network includes a gate unit and a storage unit, the long short-term memory neural network generates a vector and selects candidate information between 0 and 1, in the input gate, the vector I t is generated by the hidden state h t-1 and the input x t of the current step, and f g , f i , f c , f o , f h are different stage activation functions, w i , w g , w o are weight matrices, and b i , b g , b o are biases, and the generation process can be represented as:

[0107] I t = f i (w i x t +w i h t-1 +b i ),

[0108] The forget gate F t determines whether the state C t-1 of the previous unit should be retained, and is represented as:

[0109] F t = f g (w g x t +w g h t-1 +b g ),

[0110] Then, the input candidate information I t is generated by x t-1 and h t

[0111]

[0112] Further, the current unit state C t is obtained as:

[0113]

[0114] The output gate also generates a vector Y t :

[0115] Y t = f o (w o x t +w o h t-1 +b o ),

[0116] Finally, the hidden state h t is obtained according to Y t and C t .

[0117] h t = Y t f h (C t ),

[0118] The mean square error is used as a loss function to realize the training of the long short-term memory neural network, and the mean square error L MSE is expressed as:

[0119]

[0120] Wherein, N is the total number of samples of all categories; x t is the model input data; is the reconstructed data.

[0121] In the preferred embodiment of the online abnormality monitoring and identification method of the nuclear power circulating water pump, in the training process, the maximum mean difference MMD is calculated by using the kernel skill, and k(x, y) = (φ(x), φ(y)) H is set, and the maximum mean difference MMD is expressed as:

[0122]

[0123] The potential representation of the Gaussian distribution P z is approximated by sampling from P z and by averaging the kernel k(x, y) evaluated on all sample pairs, and the MMD on the set X is calculated in one iteration, wherein the number of X is equal to the product of the training batch size B and the data length.

[0124] In the preferred embodiment of the online abnormality monitoring and identification method of the nuclear power circulating water pump, the abnormality identification network comprises a batch normalization layer, and the activation function is set to ReLU.

[0125] In the preferred embodiment of the online abnormality monitoring and identification method of the nuclear power circulating water pump, the output result update interval is 5s.

[0126] In the preferred embodiment of the online abnormality monitoring and identification method of the nuclear power circulating water pump, the potential space d < N x T.

[0127] In the preferred embodiment of the online abnormality monitoring and identification method of the nuclear power circulating water pump, the back propagation algorithm is used to train the abnormality identification network, the loss function is calculated by using chain derivation to calculate the gradient of each weight, the weight is updated according to the gradient descent algorithm, and the cost function used by the convolutional neural network is the cross-entropy function, and the formula is as follows:

[0128]

[0129] In the formula, c represents the cost, x represents the sample, n represents the total number of samples, and a represents the model output value.

[0130] In one embodiment, the online abnormality monitoring and identification method of the nuclear power circulating water pump comprises the following steps.

[0131] Step 1: Online acquisition of multi-source time series data in the circulating water pump running process, data labeling given by an expert analysis module, data normalization processing of non-overlapping sampling with a length of 2048 is performed on the signals in sequence, and the signals are randomly divided into training set or test set samples.

[0132] Step 2: In the early stage of device operation, when only healthy samples can be collected, an unsupervised abnormality monitoring module of the long-short time self-encoder with feature layer constraint is constructed. The attention mechanism and the gate structure can fully mine the state information and potential correlation characteristics of the data, and realize the state monitoring under the healthy samples.

[0133] Step 3: With the emergence of abnormal data in the monitoring data, based on the lifelong learning architecture, the existing healthy samples and the abnormal samples generated at any time are fully utilized, the full utilization of the data and the online learning and application of the model are realized. Among them, the distillation loss and the ternary loss as the loss function compensation item can alleviate the forgetting of the model to the old categories and standardize the sample distribution; the use of the example space alleviates the storage pressure of the data acquisition terminal.

[0134] Step 4: The online monitoring and identification of known / unknown abnormal states are realized by using the adjacent mean classifier.

[0135] Online acquisition of multi-source time series data in the circulating water pump running process, data labeling given by an expert analysis module, data normalization processing of non-overlapping sampling with a length of 2048 is performed on the signals in sequence, and the signals are randomly divided into training set or test set samples, specifically:

[0136] (1) Design experiments and collect data for possible abnormalities that may occur during the operation of the circulating water pump. At the same time, considering the limitations of single-source data on the overall state characterization of the circulating water pump, multiple-source signals are selected as model inputs based on analysis. Here, multi-source high-frequency data are used as inputs, and the selection of measuring points includes Z-direction vibration acceleration of the motor base, Z-direction vibration acceleration signal of the gear box, Z-direction vibration acceleration of the pump guide bearing, sound pressure near the gear box output shaft, X-direction displacement of the gear box output shaft, and Y-direction displacement of the gear box output shaft, totaling 6 time series signals, which can effectively reflect the operating state of key components such as the motor, gear box, gear box output shaft, gear box oil pump, pump guide bearing, sliding rotor, and impeller.

[0137] (2) Based on the collected data, experts perform manual labeling work, and assign labels to the data according to the actual state of operation.

[0138] (3) Non-overlapping sampling is performed on the data, and the length of the segmented single data is 2048, and the channel is 6 channels. That is, a single data is a two-dimensional data of 6x2048.

[0139] (4) The samples x1, x2, x3, …, x n are transformed as follows:

[0140]

[0141] wherein, is the minimum data value of the corresponding sample, is the maximum data value in the corresponding sample. Compared with the traditional deviation standardization, the above method can make the relevant signal features (such as spectral peak value and frequency energy distribution) unchanged, and normalize each sample in the interval [0, 1], reducing the extreme value between samples and effectively solving the problem of convergence difficulty of parallel network model, improving the training efficiency. Step 2 describes the construction of an unsupervised anomaly monitoring module of the long short-term self-encoder with feature layer constraints in the early stage of the operation of the nuclear circulating water pump when only healthy samples can be collected. The attention mechanism and the gate structure can fully exploit the state information and potential correlation characteristics of the data, and realize state monitoring under healthy samples, which is specifically:

[0142] (1) Constructing a long short-term self-encoder

[0143] The long short-term memory neural network (LSTM) overcomes the problem of "gradient disappearance or explosion" in the back propagation of traditional recurrent neural network (RNN). The gate unit and storage unit in its structure can effectively maintain and transmit the key features of the data.

[0144] First, LSTM generates a vector and selects candidate information between 0 and 1. In the input gate, the vector I t From the hidden state h t-1 and the input x of the current step t Generate. Let f g 、f i 、f c 、f o 、f h is the activation function at different stages, w i 、w g 、w o is the weight matrix, b i 、b g 、b o is the deviation, the generation process can be expressed as:

[0145] I t =f i (w i x t +w i h t-1 +b i )

[0146] Forget Gate F t Determines the state of the previous unit C t-1 Whether it should be retained, expressed as:

[0147] F t =f g (w g x t +w g h t-1 +b g )

[0148] Then, by x t and h t-1 Generate input candidate information

[0149]

[0150] Further, get the current cell state C t for:

[0151]

[0152] The output gate also generates a vector Y t :

[0153] Y t =f o (w o x t +w o h t-1 +bo )

[0154] Finally, according to Y t and C t , the hidden state h t is obtained:

[0155] h t = Y t f h (C t )

[0156] The gate unit and the storage unit in the LSTM structure can effectively maintain and transmit the key features of the data,

[0157] The LSTM-AE combines the LSTM network with the auto-encoder, which combines the powerful time series data feature mining ability of the LSTM and the data reconstruction ability of the auto-encoder, and can mine the spatio-temporal correlation features of the data in an unsupervised setting, and realize the reconstruction of the multi-source time series data. The LSTM-AE realizes the compression and reconstruction of the input signal x t by using the encoder E and the decoder D. Further, the mean square error is used as the loss function to realize the training of the model, and the mean square error L MSE can be expressed as:

[0158]

[0159] Wherein, N is the total number of samples of all categories; x t is the input data of the model; is the reconstructed data.

[0160] (2) Design feature layer constraint

[0161] In order to constrain the data distribution of the feature layer and suppress the overfitting phenomenon of the model, the maximum mean discrepancy (MMD) is used as a regularization term to align the feature distribution of the encoding layer and the Gaussian distribution. MMD measures the difference between feature distributions by mapping samples to Hilbert space. Specifically, given the latent representation Where d is the latent space (usually d z is selected as the target distribution in the reproducing kernel Hilbert space H. MMD is calculated as follows:

[0162]

[0163] Here we have the distribution Z of the low-dimensional representation and the target distribution z (i) ~ P zMMD is defined as a characteristic map φ: X→ H, where H is a reproducing kernel Hilbert space.

[0164] During training, the kernel trick can be used to compute MMD. H An alternative representation of MMD can be derived as follows:

[0165]

[0166] Here, the latent representation of the Gaussian distribution P z is approximated by sampling from P z and by averaging the kernel k(x, y) evaluated on all pairs of samples. Typically, neural network training is performed in batches, which means that the model is trained on a sub-sample of the data in one iteration. Therefore, MMD needs to be computed on a set X in one iteration, where the number of X is equal to the product of the training batch size B and the length of the data.

[0167] Based on the above (1) and (2), the long short self-encoder based on the feature layer constraint can be applied as an abnormality monitoring module in the case of no abnormal sample.

[0168] Step 3: With the occurrence of abnormal data in the monitoring data, the life-long learning architecture is used to make full use of the existing healthy samples and the abnormal samples generated at any time, so that the data is fully utilized and the model is learned and applied online. Among them, the distillation loss and the triplet loss as the loss function compensation item can alleviate the forgetting of the model to the old category and regulate the sample distribution; the use of the example space alleviates the storage pressure of the data acquisition terminal. Specifically:

[0169] (1) Feature layer migration module

[0170] The long short self-encoder module trained based on the healthy samples has learned certain feature extraction ability and feature fusion ability. Here, the encoder of the long short self-encoder is not retrained as a feature extractor of the anomaly recognition network. In the subsequent training, in order to ensure the stability of the feature extractor, the learning rate of the feature extractor is set to one tenth of that of other modules.

[0171] (2) Network training module with fusion of triplet loss and distillation loss

[0172] To realize online learning based on online acquisition of "data flow", the application proposes a life-long learning training method. In life-long learning training, first, the new samples and a small amount of stored old sample examples need to be merged; second, the classification network is trained with the merged data set, and the model parameters are updated by reducing the network loss. The loss function includes three parts: cross-entropy loss, distillation loss and triplet loss, and the functions are as follows:

[0173] Cross-entropy L CE It is widely used in classification problems to improve the classification accuracy of the model on new samples.

[0174]

[0175] Where s is the sample label of the new class; g y is the classifier; δ is the discriminant function, which outputs 1 if and only if y=y i , otherwise 0. Distillation loss L DL is used to reproduce the output of the old sample in the original model in the new model to suppress the forgetting of learned knowledge.

[0176]

[0177] Where q i is the prediction output of the original model on the old sample; y is the learned class.

[0178] Triplet loss is an effective method to measure meta-learning, which aims to learn the mapping from the original feature space to the low-dimensional dense embedding space, so that samples of the same class are closer and samples of different classes are farther apart, thereby realizing the constraint on the feature distribution of the data.

[0179] It is composed of three neural networks sharing parameters in parallel (its essence is the same neural network), where x1 of c classes is called anchor, x2 from c classes is called positive sample, and x3 from other classes is called negative sample. The loss function Loss TN of the triplet network is as follows:

[0180] L TN = max(0, ||N(x1)-N(x2)|| 2 -||N(x1)-N(x3)|| 2 +m)

[0181] Where N(x i ) is the output of the neural network for sample x i ; m is the minimum boundary distance set.

[0182] (3) Bounded example space management module

[0183] The method uses bounded examples to realize the relearning of old data. Without loss of generality, it is assumed that the stored old examples are limited to a fixed number. Before the next online training, it is necessary to determine the number of samples that can be retained for healthy samples and known abnormal samples. Assuming that the total amount of samples that can be stored by the hardware device is K, and the total number of current new healthy state samples and old state sample categories is t, then the number of samples that can be retained by each sample is m=K / t (rounded down). In order to select representative samples for storage (examples), it is necessary to sort the importance of new samples, and select the top m most important new healthy state samples as examples. In addition, it is also necessary to reduce the number of old healthy state examples to m.

[0184] The new sample storage management of the bounded example space is shown in Algorithm 1, and the old sample reduction management of the bounded example space is shown in Algorithm 2

[0185] Algorithm 1 New sample storage management algorithm of bounded example space

[0186]

[0187] Algorithm 2 New sample storage management algorithm of bounded example space

[0188]

[0189] Step 4 uses the nearest mean classifier to realize online monitoring and identification of known / unknown abnormal states. Specifically:

[0190] (1) Known abnormality identification based on the nearest mean classifier

[0191] The present application measures the data category from the perspective of feature distribution, that is, the data closest to the prototype data of a certain category is the data of that category, and a known abnormality identification algorithm is designed based on this. Compared with the traditional deep model which directly uses a fully connected neural network for classification, the weight of the neural network will have a more serious forgetting problem than the feature layer.

[0192] Calculate the average feature vector of each category: calculate the average feature vector of the yth health state sample feature output by the neural network:

[0193]

[0194] Where, P y is the example set of y.

[0195] The Euclidean distance is used to distinguish the sample categories, and the sample to be classified is classified into the category closest to the average feature vector:

[0196]

[0197] Based on the result of anomaly identification, maintenance personnel can prioritize the predicted source of anomaly for repair. Determining the predicted source of anomaly is normal before carrying out the regular repair process can effectively reduce the workload of operation and maintenance personnel.

[0198] (2) Abnormal monitoring based on reconstruction error

[0199] Based on the reconstruction error of the health sample, the abnormal monitoring threshold is set. When the reconstruction error is higher than the threshold, the data is identified as abnormal data.

[0200] In step 4, the batch normalization layer is added to the anomaly identification model to improve the training speed and model generalization ability, and the activation function is set to ReLU. The back propagation algorithm is used to train the fault diagnosis model, the gradient of each weight is calculated using the chain rule, and the weight is updated according to the gradient descent algorithm. The cost function used in the convolutional neural network is the cross-entropy function, and its formula is:

[0201]

[0202] Where c represents the cost, x represents the sample, n represents the total number of samples, and a represents the model output value.

[0203] Softmax function: Normalized exponential function, which is a generalization of the logistic function, is defined as follows:

[0204]

[0205] Where V i is the output of the output unit before the classifier. i represents the class index, and the total number of classes is c. S i represents the ratio of the current element exponent to the sum of all element exponents. It can "compress" a K-dimensional vector containing any real numbers into another K-dimensional real vector, so that the range of each element is between 0 and 1, and the sum of all elements is 1. That is: through the softmax exponent, the output value of multi-classification can be converted into a relative probability.

[0206] The method for online abnormality monitoring and identification of a nuclear power circulating water pump trains the model incrementally in an industrial actual manner of continuously obtaining new abnormality samples, and does not need to retrain by using an offline method when a new data type appears. In an engineering application stage, the monitoring data are input into multi-source time series data, the model outputs monitoring results and identification results based on known abnormalities, and an output result update interval is set to 5 s. Maintenance personnel repair the abnormality source predicted by the model in priority. When it is determined that the predicted abnormality source is normal, the normal maintenance process is carried out, so that the workload of the maintenance personnel is effectively reduced. The nuclear power circulating water pump test bench is used to obtain the circulating water pump state monitoring data to verify the method. Figures 2-5 As shown in FIG. 8, the test bench mainly includes a motor, a planetary gear box, a guide bearing, an impeller and the like. The test bench data acquisition system includes a vibration acceleration sensor (PCB 352C33), a sound pressure sensor (SKC-MNP21), an axial displacement sensor (MIRADT3005), a temperature sensor (PT 100) and a data acquisition instrument (Yiheng-8018) and the like, a sampling frequency is set to 10240, a sampling point number is 32768, and a sampling interval is 60 s.

[0207] The present application carries out experimental research and method verification based on the circulating water pump test bench. The experiment is designed and data are collected in view of possible abnormalities occurring in the circulating water pump operation, as shown in Table 3.

[0208] Table 3 Circulating water pump health monitoring experiment

[0209]

[0210] Meanwhile, considering the limitation of single-source data on the overall state representation of the circulating water pump, multi-source signals are selected as the model input according to the analysis, as shown in FIG. 9. Figure 3 Here, multi-source high-frequency data are used as the input, and the selection of the measuring points includes a motor base Z-direction vibration acceleration, a gear box body Z-direction vibration acceleration signal, a pump guide bearing Z-direction vibration acceleration, a gear box output shaft nearby sound pressure, a gear box output shaft X-direction displacement and a gear box output shaft Y-direction displacement, a total of 6 time series signals, which can effectively reflect the operation states of the motor, the gear box, the gear box output shaft, the gear box oil supply pump, the pump guide bearing, the sliding rotor and the impeller and the like. The online abnormality state identification method for the circulating water pump fusing multi-source time series data proposed in the present application can collect data in real time during the equipment operation, and once the fault data are collected, the model can be trained online, the online identification of the abnormality state of the nuclear power circulating water pump and the online detection of unknown abnormalities are realized, and the method has a good engineering application prospect.

[0211] The long short-term memory autoencoder with feature layer constraint is trained based on multi-source health data to learn the feature expression ability of multi-source heterogeneous data in an unsupervised manner, and a loss function is as shown in Figure 3 .

[0212] It can be seen that the long short-term self-encoder with feature distribution constraint converges to the final solution at a faster speed, and the long short-term self-encoder without feature layer constraint can also converge to the optimal solution but the convergence speed is obviously slower than the proposed method, and the final mean square error values are relatively close. The self-encoder performs the worst, and the results prove that its convergence speed is slower than the above two methods, and the final mean square error is larger.

[0213] Further, the self-encoder, the long short-term self-encoder and the long short-term self-encoder with feature distribution constraint are used to reconstruct health samples and abnormal samples in test data of different categories, and a reconstruction error distribution histogram is as shown in Figure 4 , Figure 5 , Figure 6 From the perspective of data probability density distribution, the probability density distribution of the reconstruction error of the health samples of the proposed method is more concentrated, and the reconstruction effect of the health samples is more stable. At the same time, the constraint of the feature distribution of the proposed method is conducive to inhibiting the generalization phenomenon of the model and enhancing the sensitivity to abnormal data. Therefore, the distinguishability of the reconstruction error of the proposed method is better than that of the long short-term self-encoder.

[0214] The monitoring results of different abnormalities are as shown in Figure 7 It can be seen that the monitoring results of the proposed method for each type of abnormality are good, and the reliability in application is high. The proposed method is compared with iCaRL and fine-tuning method with the same network structure and the same hyperparameters, and the results are as shown in Figure 8 Compared with the proposed method, iCaRL and fine-tuning method perform worse. Among them, the fine-tuning method does not have knowledge replay and produces “catastrophic forgetting”.

[0215] The confusion matrix is drawn for the online identification result of the abnormality. The row of the confusion matrix represents the true value, and the column of the confusion matrix represents the predicted value, as shown in Figure 9 From the above results, it can be seen that after increasing the abnormality type, the proposed method can maintain an evaluation accuracy of more than 92% in the online identification task of abnormal state, indicating the superiority of the method.

[0216] Finally, it should be noted that the described embodiments are only part of the embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor belong to the scope of protection of the present application.

[0217] The foregoing merely illustrates some exemplary embodiments of the application, and no doubt numerous modifications and alterations thereto will be apparent to those skilled in the art. Accordingly, the above description is intended for purposes of illustration only and should not be construed as limiting the scope of the application.

Claims

1. A method for online abnormal monitoring and identification of a nuclear power circulating water pump, characterized in that, It comprises the following steps: The first step is to obtain multi-source time series data in the operation process of the nuclear circulating water pump online and give data labels, and then perform non-overlapping sampling of the multi-source time series data in a predetermined length, and then divide the multi-source time series data into training set samples or test set samples; In the second step, the healthy samples are collected in the early stage of the nuclear power circulating water pump operation, and the long-short time autoencoder is used to realize the state monitoring of the healthy samples. The maximum mean difference is used as a regularization term to align the feature distribution of the encoding layer and the Gaussian distribution. Given the latent representation where d is the latent space, h represents all time steps in one iteration, The corresponding features of h sub-samples are selected in the reproducing kernel Hilbert space The Gaussian distribution is selected as the target distribution, and the maximum mean difference is calculated as follows ， wherein, is a feature of a sample x i ; is a sampling result with a target distribution being a Gaussian distribution; represents a mapping from a sample feature space to a reproducing kernel Hilbert space ; low-dimensional representation of the distribution and set of target distributions , the maximum mean discrepancy is defined as a feature map where H is a reproducing kernel hilbert space, the maximum mean discrepancy over the set X is computed in one iteration, where the number of X is equal to the product of the training batch size B and the data length; The third step is to monitor the occurrence of abnormal data in the operation process of the nuclear circulating water pump, construct an anomaly recognition network based on the encoder of the long short-term self-encoder as a feature extractor and a fully connected neural network as a classifier, construct a joint loss function based on cross entropy, distillation loss and triplet loss to retain the memory of old data when learning new data, and construct the encoder of the long short-term self-encoder as a feature extractor of the anomaly recognition network in the joint loss function; Cross-entropy L CE is: ， where x i is the sample input; y is the category; y i is the predicted label of the i-th sample; s is the sample label of the new category; log is the exponential function; g y is the classifier; is the discriminant function, outputting 1 if and only if y=y i , otherwise 0; Distillation loss L DL reproducing the output of old samples in the new model on the original model to suppress forgetting of learned knowledge: ， Among them, q i is the prediction output of the original model for the old sample; s-1 represents the number of learned categories; Triplet loss function is: ， wherein, is the output of the neural network for the sample x i ; m is the set minimum boundary distance, C is the class is called an anchor, is called a positive sample from class C, is called a negative sample from other classes, respectively. In the fourth step, the real-time data of the nuclear power circulating water pump is monitored and identified on line by using the adjacent mean value classifier. The average feature vector μ of the yth health state sample characteristics output by the abnormal identification network is calculated. y : ， where P y is a set of examples of class y; is a mapping function of the feature extractor; m is the number of samples of class y; The Euclidean distance is used to distinguish the sample categories, and the sample to be classified is classified into the category closest to the average feature vector: ， where t is the total number of classes learned in the final learning; is the predicted value for the class. The abnormal monitoring threshold is set based on the reconstruction error of the health sample, and the abnormal data is determined when the reconstruction error is higher than the abnormal monitoring threshold. The multi-source time series data includes motor base Z-direction vibration acceleration, gear box body Z-direction vibration acceleration signal, pump guide bearing Z-direction vibration acceleration, gear box output shaft nearby sound pressure, gear box output shaft X-direction displacement and gear box output shaft Y-direction displacement. The long short-term memory neural network includes a gate unit and a storage unit. The long short-term memory neural network generates a vector and selects candidate information between 0 and 1. In the input gate, the vector I t is generated by the hidden state h t-1 and the input x t of the current step, and f g , f i , f c , f o , f h are different stage activation functions, w i , w g , w o are weight matrices, and b i , b g , b o are biases. The generation process is represented as: ， forget gate F t determines the state C of the preceding cell t-1 whether it should be retained, denoted by: ， Then, by x t and h t-1 generating input candidate information : ， Further, the current cell state C is obtained t is: ， The output gate also generates a vector Y t : ， Finally, the hidden state h t and C t is obtained as t : ， The mean square error is used as a loss function to train the long short-term memory neural network, and the mean square error L MSE is represented as: ， Wherein, N is the total number of samples of all categories; Input data for the model; The reconstructed data, the LSTM-AE combines the long short-term memory neural network with the auto-encoder, and the LSTM-AE adopts the composition of the encoder E and the decoder D to realize the compression and reconstruction of the input signal.

2. The method according to claim 1, characterized in that, The predetermined length is 2048, the length of the segmented single data is 2048, the channel is 6 channels, and the single data is 6x2048 two-dimensional data.

3. The method according to claim 1, characterized in that, for a sample in the training set samples or the test set samples the following transformation is performed: ， wherein, is the minimum data value of the corresponding sample, is the maximum data value in the corresponding sample, each sample is normalized in the interval [0,1].

4. The method according to claim 1, characterized in that, During training, the kernel trick is used to compute the maximum mean discrepancy, denoted as , where the maximum mean discrepancy is expressed as: ， Gaussian distribution P z The potential representation is approximated by sampling from P z MMD over the set X is computed in one iteration, where the number of X is equal to the product of the training batch size B and the data length.

5. The method according to claim 1, wherein, The anomaly recognition network contains a batch normalization layer, and the activation function is set to ReLU.

6. The method according to claim 1, wherein, The output result update interval is 5s.

7. The method according to claim 1, wherein, latent space .

8. The method according to claim 1, wherein, The back propagation algorithm is used to train the anomaly recognition network, the loss function is used to calculate the gradient of each weight, the weight is updated according to the gradient descent algorithm, and the cost function used by the convolutional neural network is the cross entropy function, and its formula is: ， Wherein, C represents the cost, x represents the sample, n represents the total number of samples, and a represents the model output value.

Images