Industrial equipment fault detection method with time series state variables

Abstract

A method for fault detection of industrial equipment with time series state variables belongs to the field of industrial equipment fault detection. The present invention aims at the problem that the fault detection of existing industrial systems relies on fault data to establish a detection model, and the difficulty in obtaining fault data makes the method difficult to apply. It includes: collecting time series state data as training data; calculating the mean and standard deviation of the m-dimensional feature variables of the training data, and standardizing the training data; using a sliding window to segment the standardized time series data to obtain sequence data P, and training the LSTM autoencoder network; then obtaining the error sequence of the training data based on the trained LSTM autoencoder network; analyzing the error sequence based on the method of extreme value theory to obtain a fault threshold; then adjusting the fault threshold based on the sequence data of the equipment to be detected to obtain the adjusted threshold, thereby realizing fault detection of industrial equipment. The present invention is used for fault detection of industrial equipment whose variables have time series characteristics.

Description

Technical Field

[0001] The invention relates to an industrial equipment fault detection method with time series state variables, and belongs to the field of industrial equipment fault detection. Background Art

[0002] Fault detection has a wide range of applications and an important position in the industrial field. Industrial systems and equipment usually have extremely complex components, and their failures will cause huge harm and losses, so it is of great significance to monitor and detect faults.

[0003] At present, it is extremely difficult to detect faults in industrial equipment. On the one hand, the industrial system is large and complex in structure, and the parameter dimension of the monitoring data is high. The intrinsic characteristics of the system are contained in complex parameter relationships, so it is difficult to extract features that can reflect its intrinsic performance using traditional feature extraction methods. On the other hand, the failure rate of industrial systems is low in practice, and it is very costly and difficult to collect fault data. Therefore, it is difficult to apply the method of analyzing fault data and establishing fault detection.

[0004] Therefore, in fault detection in the industrial field, the model needs to have strong feature extraction capabilities and be able to establish and complete detection without the need for fault data. Summary of the invention

[0005] In view of the problem that fault detection of existing industrial systems relies on fault data to establish a detection model, and the difficulty in acquiring fault data makes the method difficult to apply, the present invention provides an industrial equipment fault detection method with time series state variables.

[0006] The present invention provides an industrial equipment fault detection method with time series state variables, comprising:

[0007] The time series state data of n moments when the industrial equipment is operating normally is collected with a set step length as training data. The training data at each moment includes m-dimensional feature variables, where m is the total number of feature variables of the industrial equipment.

[0008] Calculate the mean μ and standard deviation σ of the m-dimensional feature variables of the training data, and standardize all the training data at each moment to obtain the standardized time series data;

[0009] The standardized time series data is segmented using a sliding window with a preset length of L to obtain n-L+1 sequence data P with a length of L;

[0010] The sequence data P is used to train the LSTM autoencoder network to obtain sequence output data; the sequence output data has the same dimension as the sequence data P; the mean square error is calculated from the sequence output data and the sequence data P, and the mean square error is used as the loss to optimize the network parameters. After reaching the preset number of training times, the trained LSTM autoencoder network is obtained;

[0011] Then, the sequence data P is input into the trained LSTM autoencoder network to obtain the output data, and the error sequence is calculated based on the output data and the sequence data P;

[0012] The error sequence is analyzed based on the extreme value theory method to obtain the fault threshold;

[0013] Input the sequence data of the device to be detected into the trained LSTM autoencoder network to obtain output sequence data; determine the adjustment coefficient of the fault threshold according to the output sequence data and the sequence data of the device to be detected, adjust the fault threshold, and obtain the adjusted threshold;

[0014] The square of the difference between the sequence data of the device to be detected and the output sequence data is compared with the adjusted threshold value. If the square of the difference is greater than or equal to the adjusted threshold value, the device to be detected is determined to be in a fault state; otherwise, it is in a normal state.

[0015] According to the industrial equipment fault detection method with time series state variables of the present invention, the training data includes Tennessee-Eastman industrial process data.

[0016] According to the industrial equipment fault detection method with time series state variables of the present invention, the training data is represented by X train :X train ={X (1) ,X (2) ,…,X (n)},

[0017] Where X (i) ∈R m , i=1, 2, 3,..., n; R m Represents X (i) It has m-dimensional characteristics;

[0018] μ={μ1,μ2,…,μ m},μ j is the mean of the feature variable of the jth dimension, j∈[1,m],

[0019] σ={σ1,σ2,…,σ m},σ j is the standard deviation of the feature variable of the jth dimension.

[0020] According to the industrial equipment fault detection method with time series state variables of the present invention, the method of performing standardization processing on all training data at each moment to obtain standardized time series data includes:

[0021]

[0022] In the formula is the j-th dimension feature variable of the training data at the i-th moment; It is the standardized data of the j-th dimension feature variable of the training data at the i-th moment; all the standardized data constitute the standardized time series data;

[0023] The sequence data P of length L is:

[0024] p={P1,P2,P3,…,P n-L+1}, i1∈[1,n-L+1], Indicates P i1 The length is L and has m-dimensional features.

[0025] According to the industrial equipment fault detection method with time series state variables of the present invention, P i1 After input training, the LSTM autoencoder network obtains the output data P′ i1 , by P i1 and P′ i1 Calculate the error e i1 :

[0026] e il =(P i1 -P′ i1 ) 2 ,

[0027] All errors e i1 Composition error sequence E:

[0028] E={e1,e2,…,e n-L+1}.

[0029] According to the industrial equipment fault detection method with time series state variables of the present invention, the error sequence E is analyzed based on the extreme value theory method to obtain the fault threshold

[0030] Adjusted Threshold for:

[0031]

[0032] Where ε is the adjustment coefficient:

[0033]

[0034] Where P test is the serial data of the device to be tested, P′ test Output sequence data.

[0035] According to the industrial equipment fault detection method with time series state variables of the present invention, if If the detected device is in fault state, it is in normal state otherwise.

[0036] According to the industrial equipment fault detection method with time series state variables of the present invention, the process of using sequence data P to train the LSTM autoencoder network and obtain sequence output data includes:

[0037] The LSTM autoencoder network includes an input fully connected layer, an LSTM encoder, an LSTM decoder, and an output fully connected layer;

[0038] The sequence data P is input into the fully connected layer to obtain the first-level input sequence data;

[0039] The first-level input sequence data is input into the LSTM encoder to obtain the representation of the hidden state of the first-level input sequence data; the representation of the hidden state of the first-level input sequence data and the first-level input sequence data are then input into the LSTM decoder to obtain the reconstructed data of the first-level input sequence data, and the reconstructed data is then passed through the output fully connected layer to obtain the sequence output data.

[0040] According to the industrial equipment fault detection method with time series state variables of the present invention, the sequence data P = {P1, P2, P3, ..., P n-L+1}After inputting the fully connected layer, the first-level input sequence data X is obtained E :

[0041] In the formula With P i1 correspond;

[0042] The representation of the hidden state of the first-level input sequence data h E for:

[0043]

[0044] The reconstructed data of the first-level input sequence data is h D for

[0045]

[0046] remove Other Depend on and Input the LSTM decoder together to get;

[0047] Reconstruct the data into h D The sequence output data is obtained through the output fully connected layer.

[0048] According to the industrial equipment fault detection method with time series state variables of the present invention, the error sequence E is analyzed based on the method of extreme value theory, including:

[0049] Analyze the error sequence E according to the data extreme value threshold T:

[0050]

[0051] t is the high quantile threshold, is the first coefficient obtained by fitting GPD using the maximum likelihood estimation method, is the first coefficient obtained by fitting GPD using the maximum likelihood estimation method, q is the probability value, N t It is the number of data in the error sequence E whose value is greater than the high percentile threshold t.

[0052] Beneficial effects of the present invention: The method of the present invention is implemented based on an autoencoding neural network and extreme value theory, can effectively detect faults in industrial systems, and has the characteristics of adaptive dynamic adjustment in actual detection.

[0053] The method of the present invention can effectively extract the coupling characteristics of the timing information and parameters in the data, has no specific requirements for the distribution of the error sequence, and has strong universality. In addition, the dynamic adjustment function of the fault threshold designed in the detection method of the present invention can continuously adjust the threshold in actual applications, and has stronger adaptive ability.

[0054] The method of the present invention can be carried out by using normal state data when training the network model, and has higher practicality in fault detection in the industrial field. At the same time, the present invention can also be used for fault detection in other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flowchart of using sequence data P to train the LSTM autoencoder network to obtain sequence output data;

[0056] Figure 2 This is a schematic diagram of the internal structure of an LSTM encoder or LSTM decoder. The internal structure of the LSTM encoder and LSTM decoder is the same. t ,I t , O t They represent the forget gate, input gate, and output gate, which are the three control units in LSTM. σ and tanh are activation functions in the neural network, representing the Sigmoid function and the hyperbolic tangent function, respectively. Mt is the temporary memory information calculated by the tanh function. tRepresents the input at the current moment, h t-1 Represents the input of the neuron at the previous moment; C t-1 Represents the output of the memory unit at the previous moment, C t Represents the memory unit output at the current moment, h t Represents the output of the current neural unit;

[0057] Figure 3 It is a schematic diagram of segmenting the standardized time series data using a sliding window of preset length L to obtain n-L+1 sequence data P of length L;

[0058] Figure 4 It is a schematic diagram of the LSTM autoencoder network processing sequence data p;

[0059] Figure 5 is the adjustment coefficient ε as Schematic diagram of the changing curve, in which e=(p test -P′ test ) 2 . DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0061] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0062] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0063] Specific implementation method 1. Combination Figures 1 to 4 As shown, the present invention provides an industrial equipment fault detection method with time series state variables, comprising:

[0064] The time series state data of n moments when the industrial equipment is operating normally is collected with a set step length as training data. The training data at each moment includes m-dimensional feature variables, where m is the total number of feature variables of the industrial equipment.

[0065] Calculate the mean μ and standard deviation σ of the m-dimensional feature variables of the training data, and standardize all the training data at each moment to obtain the standardized time series data;

[0066] The standardized time series data is segmented using a sliding window with a preset length of L to obtain n-L+1 sequence data P with a length of L;

[0067] The sequence data P is used to train the LSTM autoencoder network to obtain sequence output data; the sequence output data has the same dimension as the sequence data P; the mean square error is calculated from the sequence output data and the sequence data P, and the mean square error (MSE) is used as the loss to optimize the network parameters. After reaching the preset number of training times, the trained LSTM autoencoder network is obtained;

[0068] Then, the sequence data P is input into the trained LSTM autoencoder network to obtain the output data, and the error sequence is calculated based on the output data and the sequence data P;

[0069] The error sequence is analyzed based on the extreme value theory method to obtain the fault threshold;

[0070] Input the sequence data of the device to be tested into the trained LSTM autoencoder network to obtain the output sequence data; determine the adjustment coefficient of the fault threshold according to the output sequence data and the sequence data of the device to be tested, adjust the fault threshold, and obtain the adjusted threshold; for the sample to be tested, Based on this, a specific dynamic adjustment function is designed to adjust the error threshold Adaptive adjustment of fault threshold The new threshold is obtained to detect faults in the current data to be tested, and better results can be obtained;

[0071] The square of the difference between the sequence data of the device to be detected and the output sequence data is compared with the adjusted threshold value. If the square of the difference is greater than or equal to the adjusted threshold value, the device to be detected is determined to be in a fault state; otherwise, it is in a normal state.

[0072] The LSTM autoencoder network is established based on a long short-term memory network (Long-Short Term Memory, LSTM), and the input dimension and output dimension of the network are the same.

[0073] As an example, the training data includes Tennessee-Eastman Industrial Process (TEP) data.

[0074] This implementation is for fault detection in the industrial field. Model training can be completed with only normal data. The analysis method based on extreme value theory has no requirements on the distribution of error sequences and has wide adaptability. The designed dynamic threshold adjustment function can adaptively adjust the fault threshold to improve the fault detection results.

[0075] Furthermore, the training data is represented as X train :X train ={X (1) ,X(2) ,…,X (n)}, X traiN is time series data;

[0076] Where X (i) ∈R m , i=1, 2, 3,..., n; R m Represents X (i) It has m-dimensional characteristics;

[0077] μ={μ1,μ2,…,μ m},μ j is the mean of the feature variable of the jth dimension, j∈[1,m],

[0078] σ={σ1,σ2,…,σ m}, ρ j is the standard deviation of the feature variable of the jth dimension.

[0079] Furthermore, the method of normalizing all training data at each moment to obtain standardized time series data includes:

[0080]

[0081] In the formula is the j-th dimension feature variable of the training data at the i-th moment; It is the standardized data of the j-th dimension feature variable of the training data at the i-th moment; all the standardized data constitute the standardized time series data;

[0082] The sequence data P of length L is:

[0083] P={P1,P2,P3,…,P n-L+1}, i1∈[1,n-L+1], Indicates P i1 The length is L and has m-dimensional features.

[0084] The preset length of the sliding window L can be selected as 1. The standardized time series data contains data at n time points, and the data at each time point contains m features. From time point 1 to n, the data is continuously segmented using a time window of size L to obtain sequence data P.

[0085] Going further, P i1 After input training, the LSTM autoencoder network obtains the output data P′ i1 , by P i1 and P′ i1 Calculate the error e i1 :

[0086] ei1 =(P i1 -P′ i1 ) 2 ,

[0087] All errors e i1 Composition error sequence E:

[0088] E={e1,e2,…,e n-L+1}.

[0089] Based on the extreme value theory method, by setting parameters t and q, the error sequence E is analyzed to obtain the fault threshold

[0090] In order to make the model have strong adaptive ability, in the actual test process, the threshold is adjusted according to the current input data. Make fine adjustments.

[0091] Adjusted Threshold for:

[0092]

[0093] Where ε is the adjustment coefficient:

[0094]

[0095] Where P test is the serial data of the device to be tested, P′ test Output sequence data.

[0096] like If the detected device is in fault state, it is in normal state otherwise.

[0097] Figure 5 As shown, the adjustment coefficient ε increases with Its advantage is that it can cover a wider range of normal data without affecting fault detection, thereby reducing the false alarm rate of fault detection and achieving the purpose of reducing maintenance costs.

[0098] Going a step further, combined Figure 1 and Figure 4 As shown in the figure, the process of using sequence data P to train the LSTM autoencoder network and obtain sequence output data includes:

[0099] The LSTM autoencoder network includes an input fully connected layer, an LSTM encoder, an LSTM decoder, and an output fully connected layer;

[0100] The sequence data P first passes through the common input fully connected layer to obtain the first-level input sequence data;

[0101] The first-level input sequence data is input into the LSTM encoder to obtain the representation of the hidden state of the first-level input sequence data; the representation of the hidden state of the first-level input sequence data and the first-level input sequence data are then input into the LSTM decoder to obtain the reconstructed data of the first-level input sequence data, and the reconstructed data is then passed through the output fully connected layer to obtain the sequence output data.

[0102] In this practical method, the encoding and decoding structures are both LSTM. The data is input into the LSTM encoder to obtain the representation of the hidden state of the data. The hidden state of the encoder and the original data are input into the LSTM decoder to continuously reconstruct the original time series data. Among them, the input and output sequence lengths and feature dimensions are the same. MSE can be used to calculate the input and output errors and optimize the network parameters.

[0103] Going a step further, combined Figure 4 As shown, the sequence data P = {P1, P2, P3, ..., P n-L+1}After inputting the fully connected layer, the first-level input sequence data X is obtained E :

[0104] In the formula With P i1 correspond;

[0105] The representation of the hidden state of the first-level input sequence data h E for:

[0106]

[0107] The reconstructed data of the first-level input sequence data is h D for

[0108]

[0109] remove Other Depend on and Input the LSTM decoder together to get;

[0110] Reconstruct the data into h D The final sequence output data is obtained through the common output fully connected layer.

[0111] Where 1≤i1≤L-1.

[0112] Furthermore, the analysis of the error sequence E based on the extreme value theory method includes:

[0113] For a random variable, no matter what distribution it follows as a whole, its extreme values ​​should have the same distribution. The distribution followed by such extreme values ​​is called extreme value distribution. It is expressed as:

[0114]

[0115] Where x represents the random variable in the extreme value distribution, γ is the extreme value index, which determines the shape of the function distribution. Represents the set of real numbers.

[0116] There is also the second extreme value theorem. For the cumulative distribution function F, if and only if the σ function exists, all When the threshold t→τ, we have:

[0117]

[0118] A further understanding of the above formula is:

[0119]

[0120] Where P represents the conditional probability and X is the random variable.

[0121] That is, the part (Xt) exceeding the threshold t should follow the generalized Pareto distribution (GPD).

[0122] Then, the error sequence E is analyzed according to the data extreme value threshold T:

[0123]

[0124] t is the high quantile threshold, is the first coefficient obtained by fitting GPD using the maximum likelihood estimation method, is the first coefficient obtained by fitting GPD using the maximum likelihood estimation method, q is the probability value, N t is the number of data in the error sequence E whose value is greater than the high quantile threshold t. The high quantile threshold t and probability value q need to be set by yourself, and n is the total number of observations.

[0125] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. It should therefore be understood that many modifications may be made to the exemplary embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in a manner different from that described in the original claims. It should also be understood that the features described in conjunction with a single embodiment may be used in other described embodiments.

Claims

1. A method for detecting faults in industrial equipment with time series state variables, characterized in that include, The time series state data of n moments when the industrial equipment is operating normally is collected with a set step length as training data. The training data at each moment includes m-dimensional feature variables, where m is the total number of feature variables of the industrial equipment. Calculate the mean μ and standard deviation σ of the m-dimensional feature variables of the training data, and standardize all the training data at each moment to obtain the standardized time series data; The standardized time series data is segmented using a sliding window with a preset length of L to obtain n-L+1 sequence data P with a length of L; The sequence data P is used to train the LSTM autoencoder network to obtain sequence output data; the sequence output data has the same dimension as the sequence data P; the mean square error is calculated from the sequence output data and the sequence data P, and the mean square error is used as the loss to optimize the network parameters. After reaching the preset number of training times, the trained LSTM autoencoder network is obtained; Then, the sequence data P is input into the trained LSTM autoencoder network to obtain the output data, and the error sequence is calculated based on the output data and the sequence data P; The error sequence is analyzed based on the extreme value theory method to obtain the fault threshold; Input the sequence data of the device to be detected into the trained LSTM autoencoder network to obtain the output sequence data; Determine the adjustment coefficient of the fault threshold according to the output sequence data and the sequence data of the device to be detected, adjust the fault threshold, and obtain the adjusted threshold; The square of the difference between the sequence data of the device to be detected and the output sequence data is compared with the adjusted threshold value. If the square of the difference is greater than or equal to the adjusted threshold value, the device to be detected is determined to be in a fault state; otherwise, it is in a normal state.

2. The industrial equipment fault detection method with time series state variables according to claim 1, characterized in that: The training data consists of the Tennessee-Eastman industrial process data.

3. The industrial equipment fault detection method with time series state variables according to claim 1 or 2, characterized in that: The training data is represented as X train :X train ={X (1) , X (2) , ..., X (n) }, Where X (i) ∈R m , i=1, 2, 3,..., n; R m Represents X (i) It has m-dimensional characteristics; μ ={μ1, μ2, ..., μ m },μ j is the mean of the feature variable of the jth dimension, j∈[1,m], σ={σ1,σ2,...,σ m },σ j is the standard deviation of the feature variable of the jth dimension.

4. The industrial equipment fault detection method with time series state variables according to claim 3, characterized in that: Methods for standardizing all training data at each moment to obtain standardized time series data include: In the formula is the j-th dimension feature variable of the training data at the i-th moment; It is the standardized data of the j-th dimension feature variable of the training data at the i-th moment; all the standardized data constitute the standardized time series data; The sequence data P of length L is: i1∈[1,n-L+1], Indicates P i1 The length is L and has m-dimensional features.

5. The method for detecting industrial equipment faults with time series state variables according to claim 4, characterized in that: P i1 After input training, the LSTM autoencoder network obtains the output data P′ i1 , by P i1 and P′ i1 Calculate the error e i1 : e i1 =(P i1 -P′ i1 ) 2 , All errors e i1 Composition error sequence E: E={e1,e2,...e n-L+1 }。 6. The industrial equipment fault detection method with time series state variables according to claim 5, characterized in that: The error sequence E is analyzed based on the extreme value theory method to obtain the fault threshold Adjusted Threshold for: Where ε is the adjustment coefficient: Where P test is the serial data of the device to be tested, P′ test To output sequence data.

7. The industrial equipment fault detection method with time series state variables according to claim 6, characterized in that: like If the detected device is in fault state, it is in normal state otherwise.

8. The industrial equipment fault detection method with time series state variables according to claim 7, characterized in that: The process of using sequence data P to train the LSTM autoencoder network and obtain sequence output data includes: The LSTM autoencoder network includes an input fully connected layer, an LSTM encoder, an LSTM decoder, and an output fully connected layer; The sequence data P is input into the fully connected layer to obtain the first-level input sequence data; The first-level input sequence data is input into the LSTM encoder to obtain the representation of the hidden state of the first-level input sequence data; the representation of the hidden state of the first-level input sequence data and the first-level input sequence data are then input into the LSTM decoder to obtain the reconstructed data of the first-level input sequence data, and the reconstructed data is then passed through the output fully connected layer to obtain the sequence output data.

9. The method for detecting industrial equipment faults with time series state variables according to claim 8, characterized in that: Sequence data P = {P1, P2, P3, ..., P n-L+1 }After inputting the fully connected layer, the first-level input sequence data X is obtained E : In the formula With P i1 correspond; The representation of the hidden state of the first-level input sequence data h E for: The reconstructed data of the first-level input sequence data is h D for remove Other Depend on and Input the LSTM decoder together to get; Reconstruct the data into h D The sequence output data is obtained through the output fully connected layer.

10. The industrial equipment fault detection method with time series state variables according to claim 9, characterized in that: The analysis of the error sequence E based on the extreme value theory method includes: Analyze the error sequence E according to the data extreme value threshold T: t is the high quantile threshold, is the first coefficient obtained by fitting GPD using the maximum likelihood estimation method, is the first coefficient obtained by fitting GPD using the maximum likelihood estimation method, q is the probability value, N t It is the number of data in the error sequence E whose value is greater than the high percentile threshold t.

Images