A fault detection method based on deconvolution network and adversarial neural network

Abstract

The present invention discloses a fault detection method based on a deconvolutional neural network and an adversarial neural network. Specifically, in the absence of historical fault data, the method can detect the health status of the object described by the time series of health data using only the time series of health data. The fault detection method includes: obtaining a time series of health status from a sensor, grouping the time series using a sliding window to obtain training samples, constructing a multi-layer convolutional neural network, and training the multi-layer convolutional neural network to obtain corresponding weights. Deconvolution of the trained convolutional neural network is performed, and two sets of deep neural networks are constructed to form an adversarial neural network generator and a discriminator. The adversarial neural network is trained to obtain a health diagnostic device. The observation window data is input, and the discriminator outputs the health status probability. By introducing a deconvolutional neural network, the method solves the problem of difficulty in generating a time series corresponding to the process, improving the efficiency of time series generation. Through a game process between the data generator and the evaluator, the method ultimately obtains implicit boundary conditions for the health status. The method is not restricted by the characteristics of the fault type and can be applied to different fault types, improving the versatility of fault detection. It also does not require the addition of expensive specialized equipment, significantly reducing diagnostic costs.

Description

Technical Field

[0001] The present invention relates to a fault detection method for time series, and in particular, in the case of lack of historical fault data, health status detection of an object described by the time series can be achieved by only using the time series of health data. Background Art

[0002] Fault diagnosis is a crucial technical approach to ensuring the safe and reliable operation of process systems. With the application of data acquisition systems and intelligent sensors, vast amounts of data have been generated. Data-driven fault diagnosis methods are gaining increasing attention. Artificial intelligence (AI) approaches are a key implementation of this data-driven approach. These methods primarily utilize artificial neural networks (ANNs). Through learning, they can identify unknown fault characteristics, thus breaking away from the limitation of traditional fault diagnosis, which requires prior knowledge of fault characteristics. Learning methods are generally categorized into supervised and unsupervised learning. Supervised learning uses labeled training sets as training samples, and the trained neural network is then used for fault diagnosis. Unsupervised learning uses existing data for self-service clustering according to specific rules, automatically merging new data into the clusters. Both learning methods require a large amount of healthy and faulty data as a foundation. However, in real systems, for safety reasons, equipment is often not allowed to operate in a faulty state, resulting in missing or incomplete fault data. If the health data can cover all working conditions, the probability distribution of the health status can be directly constructed through these data. However, for health data, although the amount of health data is large, due to the limitations of working conditions, these health data cannot cover all health conditions and contain limited information. This pseudo-large sample cannot bring more information, but will aggravate the overfitting phenomenon in the training process. Therefore, it is difficult for us to directly obtain the probability density function from the sample data.

[0003] The current sample data can be regarded as a realization of the system state. Posterior probability theory can provide great help. Generative Adversarial Networks (GANs) provide a method to construct similar data through adversarial means. Taking maximum likelihood estimation as an example, its goal is to obtain the probability distribution of real data as much as possible. The generative model defines the probability distribution function P through the maximum likelihood estimation method. model (x;θ), the parameter of this distribution is θ. In the iterative calculation of the generative model, θ is continuously updated so that P model (.) The function can approximate the true probability distribution P as much as possible data (x). But the reality is that P data (x) is often unknown, so P can only be approximately simulated by sampling a large number of real data sets. data (x), and then train the probability distribution P of the generative model data(x). If the Monte Carlo method is used to continuously generate a large amount of data in the entire space, a large amount of pseudo data is generated through the interference of random noise. This generated data is mixed with the real data that can be collected. Through the game between the GAN recognizer and the generator, the effective expansion of health data is achieved. At this time, the generator has exceeded the scope of the collected health data samples and also includes health states that exist in the entire space but have not been discovered. The recognizer is diagnosing the classification of the entire space, thus breaking away from the limitations of the sample set and greatly improving the robustness of fault diagnosis. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to generate a large number of simulated time series through a deconvolutional neural network for the time series of the diagnosis object that lacks fault status due to safety reasons, and then use the game process of the recognizer and generator of the Generative Adversarial Network (GAN) to finally train a judge that can directly judge the health status of the object.

[0005] A fault detection method based on a combination of a deconvolutional network and an adversarial neural network, comprising:

[0006] Step 1: Obtain the time series of health status from the sensor , where k is the sampling time, Indicates that the time series is long enough;

[0007] Step 2: Use a sliding window of length T to divide the time series into m groups, each group includes the window time series and target state , where is the starting sampling point of the window, For time series The next state value of the window time series in ;

[0008] Step 3: Construct a multi-layer convolutional neural network CNN1. The input layer dimension of CNN1 is T and the output layer is The hidden layers include convolution layer 1, pooling layer 1, convolution layer 2, pooling layer 2, ..., convolution layer n, pooling layer n and summation layer. The specific number of hidden layers is obtained through trial and error according to different time series.

[0009] Step 4: Take the m groups of window time series obtained in step 2 and target state As a training data set, that is ( ) l =1,2,…m; Train the multi-layer convolutional neural network CNN1 to obtain the corresponding weights; the training process is described later;

[0010] Step 5: Deconvolution network DCNN1 is obtained for the trained CNN1, that is, DCNN1 has the same structure as CNN1, with opposite data flows, but the values of the corresponding weights remain unchanged, that is, the input layer of CNN1 becomes the output layer of DCNN1, and the pooling layer n, convolution layer n, ..., pooling layer 2, convolution layer 2, pooling layer 1, and convolution layer 1 of CNN1 become the convolution layer 1, pooling layer 1, convolution layer 2, pooling layer 2, ..., convolution layer n, and pooling layer n of DCNN1;

[0011] Step 6: Construct two sets of deep neural networks to form the generator G and discriminator D of the adversarial neural network, where the input of the generator G is a column vector composed of time series, and the output is the target state; the input of the discriminator D is a column vector composed of time series, and the output is the health probability;

[0012] Step 7: Train the adversarial neural network until the training is completed; the training process is described below;

[0013] Step 8: The data generator inputs new observation window data, and the judge outputs the diagnosis result.

[0014] Among them, the training process of CNN1 in step 4 is as follows:

[0015] Step 1: Calculate the forward propagation of the convolutional neural network according to the following formula:

[0016] Convolutional layer: (1)

[0017] Where, is the output of the convolutional layer, is the convolutional layer input, is the convolutional layer weight, is the convolution layer activation function threshold, is the activation function input, is the activation function,

[0018] Pooling layer: (2)

[0019] Where, is the output of the pooling layer, is the input of the pooling layer, is the maximum value of the corresponding area;

[0020] Fully connected layer: (3)

[0021] Where, is the output of the fully connected layer, is the input of the fully connected layer, is the weight of the fully connected layer, is the activation function threshold of the fully connected layer, is the activation function input, is the activation function,

[0022] Output layer: (4)

[0023] Where, is the output layer output, is the output layer weight, is the output layer activation function threshold, , for The i-th node output of , c is the number of output nodes;

[0024] Step 2: Calculate the loss function L according to formula (5)

[0025] (5)

[0026] Where, is the number of training samples, is the number of neural networks in the output layer; is the target value of the training sample; is the output of the convolutional neural network;

[0027] Step 3: Modify the weights using the gradient descent method according to formula (6)

[0028] (6)

[0029] Where, is the connection weight of the neural network, including the weights of the convolutional layer, the fully connected layer, and the output layer. is the weight of the k-th training, is the weight of the k+1th training, is the learning rate;

[0030] Step 4: Repeat steps 1 to 3 until the loss function L no longer changes.

[0031] The adversarial neural network training process in step seven is as follows:

[0032] Step 1. Randomly initialize the parameters of the generator G and the discriminator D ;

[0033] Step 2: For m input samples, calculate the risk loss function according to formula (7): ,

[0034] (7)

[0035] Where, is the loss function, , where y and are actual and estimated values, respectively; is a rule item used to prevent overfitting. ; To distinguish the risk loss function of the generator G and the discriminator D, they are respectively denoted as and ;

[0036] Step 3. Adjust the parameters of the generator G according to equations (8) and (9) until the output of the discriminator D is 1.

[0037] (8)

[0038] (9)

[0039] Where, and is the weight and threshold of generator G, and For the Iteration value, and For the Iteration value, is the learning rate;

[0040] Step 4: Fix the parameters of the generator G and adjust the parameters of the discriminator D according to equations (10) and (11).

[0041] (10)

[0042] (11)

[0043] Where, and is the weight and threshold of the discriminator D, and For the Iteration value, and For the Iteration value, is the learning rate;

[0044] Step 5: Repeat steps 2 to 4 until the risk function and All converged.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. The method for fault detection based on a deconvolutional network and a generative adversarial neural network is a data-driven intelligent method. By introducing a deconvolutional neural network, it solves the problem of difficulty in generating a time series corresponding to the process and improves the efficiency of time series generation.

[0047] 2. The present invention introduces a generative adversarial neural network. Through the game process between the data generator G and the evaluator D, the implicit boundary conditions of the health state are finally obtained. This method is not limited by the characteristics of the fault type and can be applied to different fault types, thus improving the versatility of fault detection.

[0048] 3. The present invention only requires the system's previous health data, and the data acquisition system collects current data to complete fault diagnosis. There is no need for additional expensive special equipment, thereby greatly reducing the diagnosis cost. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a structural diagram of a fault detection method based on the combination of deconvolutional network and adversarial neural network;

[0050] Figure 2 It is a schematic diagram of the relationship between convolutional neural network and deconvolutional neural network;

[0051] Figure 3 is a schematic diagram of the structure of the generator G;

[0052] Figure 4 It is a structural diagram of the discriminator D. DETAILED DESCRIPTION

[0053] The technical solution of the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments. The specific embodiments described are only used to explain the present invention and are not intended to limit the present invention.

[0054] The present invention is a fault detection method based on the combination of deconvolution network and adversarial neural network. The whole system includes a detection object, a sensor and a data processing unit. Its structure is as follows: Figure 1 shown.

[0055] The detection object includes a physical information device or system that can generate a time series, including but not limited to a production system with independent functions, mechanical equipment, electrical devices, etc.;

[0056] The sensor refers to a device that can detect physical quantities and convert them into digital quantities, including but not limited to pressure sensors, flow sensors, displacement sensors, etc.

[0057] The data processing unit refers to the combination of the hardware platform and the software that implements the algorithm. The hardware platform includes but is not limited to a microprocessor, a computer, etc.; the software refers to the software code developed based on the algorithm on the hardware platform;

[0058] The time window refers to a preset window length T, within which the sensor data values are extracted from the time series;

[0059] The convolutional neural network refers to a convolutional neural network constructed according to time window data;

[0060] The deconvolutional neural network refers to an artificial neural network with the same structure as the trained convolutional neural network, unchanged weights, but opposite data flow;

[0061] The data generator refers to a deep neural network constructed based on time window data;

[0062] The discriminator refers to a deep neural network constructed based on system state data;

[0063] The risk loss function refers to the superposition of the loss function and the rule term;

[0064] The output result refers to the probability that the trained discriminator outputs a state close to a healthy state.

[0065] The specific working steps are as follows:

[0066] Step 1: Obtain the time series of health status from the sensor , where k is the sampling time, Indicates that the time series is long enough;

[0067] Step 2: Use a sliding window of length T to divide the time series into m groups, each group includes the window time series and target state , where is the starting sampling point of the window, For time series The next state value of the window time series in ;

[0068] Step 3: Construct a multi-layer convolutional neural network CNN1. The input layer dimension of CNN1 is T and the output layer is The hidden layers include convolution layer 1, pooling layer 1, convolution layer 2, pooling layer 2, ..., convolution layer n, pooling layer n and summation layer. The specific number of hidden layers is obtained through trial and error according to different time series.

[0069] Step 4: Take the m groups of window time series obtained in step 2 and target state As a training data set, that is ( ) l =1,2,…m; Train the multi-layer convolutional neural network CNN1 to obtain the corresponding weights; the training process is described later;

[0070] Step 5: Deconvolution network DCNN1 is obtained for the trained CNN1, that is, DCNN1 has the same structure as CNN1, with opposite data flows, but the values of the corresponding weights remain unchanged, that is, the input layer of CNN1 becomes the output layer of DCNN1, and the pooling layer n, convolution layer n, ..., pooling layer 2, convolution layer 2, pooling layer 1, and convolution layer 1 of CNN1 become the convolution layer 1, pooling layer 1, convolution layer 2, pooling layer 2, ..., convolution layer n, and pooling layer n of DCNN1;

[0071] Step 6: Construct two sets of deep neural networks to form the generator G and discriminator D of the adversarial neural network, where the input of the generator G is a column vector composed of time series, and the output is the target state; the input of the discriminator D is a column vector composed of time series, and the output is the health probability;

[0072] Step 7: Train the adversarial neural network until the training is completed; the training process is described below;

[0073] Step 8: The data generator inputs new observation window data, and the judge outputs the diagnosis result.

[0074] Among them, the training process of CNN1 in step 4 is as follows:

[0075] Step 1: Calculate the forward propagation of the convolutional neural network according to the following formula:

[0076] Convolutional layer: (1)

[0077] Where, is the output of the convolutional layer, is the convolutional layer input, is the convolutional layer weight, is the convolution layer activation function threshold, is the activation function input, is the activation function, ;

[0078] Pooling layer: (2)

[0079] Where, is the output of the pooling layer, is the input of the pooling layer, is the maximum value of the corresponding area;

[0080] Fully connected layer: (3)

[0081] Where, is the output of the fully connected layer, is the input of the fully connected layer, is the weight of the fully connected layer, is the activation function threshold of the fully connected layer, is the activation function input, is the activation function, ;

[0082] Output layer: (4)

[0083] Where, is the output layer output, is the output layer weight, is the output layer activation function threshold, , for The i-th node output of , c is the number of output nodes;

[0084] Step 2: Calculate the loss function L according to formula (5)

[0085] (5)

[0086] Where, is the number of training samples, is the number of neural networks in the output layer; is the target value of the training sample; is the output of the convolutional neural network;

[0087] Step 3: Modify the weights using the gradient descent method according to formula (6)

[0088] (6)

[0089] Where, is the connection weight of the neural network, including the weights of the convolutional layer, the fully connected layer, and the output layer. is the weight of the k-th training, is the weight of the k+1th training, is the learning rate;

[0090] Step 4: Repeat steps 1 to 3 until the loss function L no longer changes.

[0091] The adversarial neural network training process in step seven is as follows:

[0092] Step 1. Randomly initialize the parameters of the generator G and the discriminator D ;

[0093] Step 2: For m input samples, calculate the risk loss function according to formula (7): ,

[0094] (7)

[0095] Where, is the loss function, , where y and are actual and estimated values, respectively; is a rule item used to prevent overfitting. ; To distinguish the risk loss function of the generator G and the discriminator D, they are respectively denoted as and ;

[0096] Step 3. Adjust the parameters of the generator G according to equations (8) and (9) until the output of the discriminator D is 1.

[0097] (8)

[0098] (9)

[0099] Where, and is the weight and threshold of generator G, and For the Iteration value, and For the Iteration value, is the learning rate;

[0100] Step 4: Fix the parameters of the generator G and adjust the parameters of the discriminator D according to equations (10) and (11).

[0101] (10)

[0102] (11)

[0103] Where, and is the weight and threshold of the discriminator D, and For the Iteration value, and For the Iteration value, is the learning rate;

[0104] Step 5: Repeat steps 2 to 4 until the risk function and All converged.

[0105] Simulation verification:

[0106] The model was validated on a 4.8MW wind turbine benchmark model. Wind data from an actual wind farm was used to activate the wind turbine model. The simulation lasted 4400s, with a sampling interval of 8ms, for a total of 440,000 data sets. The sliding window length was set to 300 sampling points. The operating system was Windows 10 and the Intel(R) core(TM) i7-10700 CPU @ 2.90GHz. Following the method recommended in this paper, the generator and identifier were basically stable after 400 steps of adversarial training. Although they could not converge to 0, they had converged to a certain range. This judge was used to judge the time series of the new observation window. We considered that four different faults were created in the blade speed time series, labeled F1, F2, F3, and F4. The starting position of the sliding window was randomly selected for each fault, resulting in 100 observation sequences. The results are shown in Table 1.

[0107] Serial number illustrate Correct number of times Number of errors Accuracy 1 No trouble 93 7 93% 2 Fault 1 (F1) 94 6 94% 3 Fault 2 (F2) 95 5 95% 4 Fault 3 (F3) 95 5 95% 5 Fault 4 (F4) 93 7 93%

[0108] From the results, it can be seen that no matter what the fault is, the accuracy rate can reach more than 93%.

[0109] Although the present invention has been described above in conjunction with the accompanying drawings, the present invention is not limited to the above-mentioned specific embodiments. The above-mentioned specific embodiments are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can make many variations without departing from the purpose of the present invention, and these are all protected by the present invention.

Claims

1. A fault detection method based on a combination of a deconvolutional network and an adversarial neural network, characterized by: Step 1: Obtain the time series of health status from the sensor Where k is the sampling time, ∞ means the time series is long enough; Step 2: Use a sliding window of length T to divide the time series into m groups, each group includes the window time series and the system state y l , where l is the starting sampling point of the window, y l For time series The next state value of the window time series, y l =x l+T+1 ; Step 3: Construct a multi-layer convolutional neural network CNN1. The input layer dimension of CNN1 is the window time series column vector N, and the output layer is the system state y l The hidden layers include convolution layer 1, pooling layer 1, convolution layer 2, pooling layer 2, ..., convolution layer n, pooling layer n and summation layer. The specific number of hidden layers is obtained through trial and error according to different time series. Step 4: Take the m groups of window time series obtained in step 2 and the target state y l As a training data set, Train the multi-layer convolutional neural network CNN1 to obtain the corresponding weights; Step 5: Deconvolution network DCNN1 is obtained for the trained CNN1, i.e., DCNN1 has the same structure as CNN1, with opposite data flows, but the values of the corresponding weights remain unchanged, i.e., the input layer of CNN1 becomes the output layer of DCNN1, and the pooling layer n, convolution layer n, ..., pooling layer 2, convolution layer 2, pooling layer 1, and convolution layer 1 of CNN1 become the convolution layer 1, pooling layer 1, convolution layer 2, pooling layer 2, ..., convolution layer n, and pooling layer n of DCNN1; the input of DCNN1 is the system state, and the output is the window time series column vector; Step 6: Randomly generate system states and input them into the deconvolution network DCNN1. The output of DCNN1 is a windowed time series column vector. Two sets of deep neural networks are constructed to form the data generator G and the discriminator D of the adversarial neural network. The input of the data generator G is the time series column vector, and the output is the inferred system state; the input of the discriminator D is the difference between the inferred system state and the actual system state, and the output result is the health probability; Step 7: Train the adversarial neural network until the training is completed; Step 8: The data generator inputs new observation window data, and the judge outputs the diagnosis result.

2. The fault detection method based on the combination of deconvolution network and adversarial neural network according to claim 1, characterized in that: In step 4, the m groups of window time series obtained in step 2 are used. With the system state y l As a training data set, Train the multi-layer convolutional neural network CNN1 and obtain the corresponding weights.

3. The fault detection method based on the combination of deconvolution network and adversarial neural network according to claim 1, characterized in that: In step 5, the deconvolution network DCNN1 is obtained for the trained CNN1.

4. The fault detection method based on the combination of deconvolution network and adversarial neural network according to claim 1, characterized in that: In step 6, two sets of deep neural networks are constructed to form the data generator G and discriminator D of the adversarial neural network, where the input of the data generator G is the time series vector and the output is the inferred system state; the input of the discriminator D is the difference between the inferred system state and the actual system state, and the output result is the health probability.

5. The fault detection method based on the combination of deconvolution network and adversarial neural network according to claim 1, characterized in that: The data generator in step 8 inputs new observation window data, and the judge outputs the diagnosis result.

Images