Instability fault monitoring method with self-learning strong generalization capability and instability fault monitoring apparatus

A computer-implemented instability fault monitoring method with self-learning strong generalization capability, characterized by comprising the following steps:

• step 1, collecting electrical data of each electronic device in a power electronic system, selecting some of the electrical data to be labeled with types of system instability faults, and combining the labeled electrical data with unlabeled electrical data to form dataset samples, wherein the specific process of collecting the electrical data is as follows:
• taking a bus of the power electronic system as a backbone, and regarding power converters in the power electronic system as module branches to be connected to the backbone, so as to construct an equivalent model of the power electronic system;
• and
• collecting electrical data from each module branch and the backbone according to a predetermined type of instability faults, which is, determining the corresponding instability fault category K to be identified, the fault problems to be solved being {stability, instability, module 1 failure, module 2 failure, . . . , module K failure};
• step 2, constructing a classification network based on a deep belief neural network framework, the classification network comprising a plurality of superimposed and connected network layers and a softmax layer connected to an end network layer, the output classification of the softmax layer connected to the end network layer being determined by a diagnostic task category K targeted, training the classification network using the dataset samples to construct a determination model for determining whether the power electronic system has an instability fault, wherein the training process comprises an unsupervised pre-training phase and a supervised back-propagation fine-tuning phase;
• in the pre-training phase, each of the network layers is trained using a contrastive divergence algorithm based on the dataset samples, so as to construct an initial determination model;
• and
• in the back-propagation fine-tuning phase, the constructed initial determination model is fine-tuned based on the labeled electrical data to obtain the determination model for determining whether the power electronic system has an instability fault;
• and
• optimizing the training process of the classification network using a reinforcement learning algorithm, wherein a reinforcement learning network takes a network structure composed of different numbers of network layers and a softmax layer as action space, and takes a loss rate of the determination model of previous and latter trainings as an optimization objective to obtain an optimization result, wherein the optimization result comprises a superimposed network structure and a corresponding predicted loss rate;
• wherein optimizing the training process comprises:
• parameters are initialized, episode=0, T is set as the maximum number of execution steps in each episode, and the optimal diagnosis rate is set;
• then, the following operations are performed in each episode for a total of T steps:
• (1) the iteration moment t=0 is initialized, and an underlying network layer structure is randomly initialized;
• (2) then, an actor selects an action at at the current moment using a ε-greedy algorithm;
• the input of the actor is a state St=DBNt at the current moment t based on the state St and following the strategy πθ, the actor takes an action at, a network layer is selected from alternative models to form the determination model at the moment t, and then a diagnostic model St+1=DBNt+1 is formed together with the corresponding softmax layer;
• (3) the optimal diagnostic accuracy is obtained by comparing the diagnostic accuracy of DBNt+1 with the diagnostic accuracy of DBNt, and at the same time, the actor obtains a reward Rt;
• (4) when the number of data pairs [St, at, Rt, St+1] stored in the buffer reaches a certain set value N (if the number does not reach N, steps (2)-(3) are repeated), training update is performed on an actor parameter, N small batch samples are mainly sampled from the buffer, and a critic is utilized to predict the corresponding action a′i that may be taken in the state Si (i=1, 2, . . . , N), as well as the obtained reward Ri, and a target return value in the state Si is calculated;
• a loss function is finally obtained:
• L
• ⁡
• (
• θ
• )
• =
• 1
• N
• ⁢
• ∑
• i
• =
• 1
• N
• (
• y
• i
• -
• Q
• ⁡
• (
• S
• i
• +
• 1
• ,
• a
• i
• ;
• θ
• )
• )
• 2
• where θ denotes a to-be-updated parameter of the actor, Si+1 denotes a state obtained after taking an actual action ai in a state Si in N samples actually collected in the buffer, and Q(Si+1, ai;
• θ) denotes an actual value in the state Si in the N samples actually collected in the buffer;
• after the loss is obtained, the actor parameter θ is updated in a gradient descent method;
• (5) a weight value of the critic is updated every C times to be consistent with a weight value of the actor, i.e., θ′=θ,
• when the execution exceeds T times, whether the recorded optimal accuracy value meets the requirement is determined;
• if the value meets the requirement, the training is terminated;
• otherwise, the next episode is started, and the neural network of the actor-critic architecture outputs a corresponding diagnostic model;
• step 3, inputting the electrical data in the power electronic system into a determination model satisfying a predetermined predicted loss rate in step 2, so as to output a determination result indicating whether the power electronic system has an instability fault, wherein the determination result comprises determining the type of instability fault, and fault diagnosis and localization of the instability source fault.