Nuclear detector fault diagnosis method and device, computer equipment and storage medium

Abstract

The application belongs to the technical field of equipment management, and discloses a nuclear detector fault diagnosis method, device, computer equipment and storage medium. The nuclear detector fault diagnosis method comprises the following steps: acquiring an output signal of the nuclear detector; inputting the output signal into a pre-trained deep belief network for fault identification to obtain a fault parameter corresponding to the output signal; and adopting a generalized likelihood ratio test classification model to perform fault classification on the fault parameter to obtain a fault category corresponding to the output signal, so that the fault of the nuclear detector can be quickly and accurately detected.

Description

Technical Field

[0001] This invention belongs to the field of equipment management technology. More specifically, this invention relates to a fault diagnosis method, apparatus, computer equipment, and storage medium for a nuclear detector. Background Technology

[0002] NaI nuclear detectors are widely used in nuclear power plant radiation protection and radioactive monitoring due to their high efficiency, high sensitivity, and stable operation at room temperature for gamma ray detection. However, NaI nuclear detectors suffer from low signal-to-noise ratio, strong background clutter, and often operate in harsh environments with high temperature and high radiation. The effects of temperature, humidity, and radiation can cause scintillator aging, photomultiplier tube malfunctions, and RC circuit abnormalities in the output analog circuit, resulting in inaccurate measurements.

[0003] Currently, in nuclear power plants, fault diagnosis of nuclear detectors mainly employs hardware redundancy. This method works by using multiple identical nuclear detectors to measure the same physical quantity; when a significant difference appears in the measurement value from a particular detector, it is determined that the detector is faulty. While this method has a simple detection principle, it requires a large number of repetitive devices, places high demands on system installation, and incurs substantial costs. Summary of the Invention

[0004] The purpose of this invention is to overcome at least one defect in the prior art and provide a method, apparatus, computer equipment, and storage medium for diagnosing nuclear detector faults that can quickly detect faults and improve fault detection efficiency.

[0005] To achieve the above-mentioned objective, the present invention provides a fault diagnosis method for a nuclear detector, comprising the following steps:

[0006] Acquire the output signal of the nuclear detector;

[0007] The output signal is input into a pre-trained deep belief network for fault identification to obtain the fault parameters corresponding to the output signal.

[0008] The fault parameters are classified using a generalized likelihood ratio test classification model to obtain the fault category corresponding to the output signal.

[0009] According to an embodiment of the fault diagnosis method for a nuclear detector of the present invention, the deep belief network includes an unsupervised network model and a backpropagation network model; the step of generating the deep belief network includes:

[0010] Acquire sample data and corresponding sample categories; the sample data includes sample fault data and sample fault-free data; the sample categories include sample fault categories and sample fault-free categories.

[0011] The sample data is input into the unsupervised network model for unsupervised training to obtain the model parameters of the unsupervised network model and the corresponding sample data features;

[0012] The sample data features and the corresponding sample fault categories are input into the backpropagation network model for supervised learning training to obtain the output results.

[0013] The model parameters of the unsupervised network model are adjusted in reverse based on the output results until training is complete.

[0014] According to an embodiment of the fault diagnosis method for a nuclear detector of the present invention, the unsupervised network model includes multiple stacked deep learning layers; the step of inputting the sample data into the unsupervised network model for unsupervised training to obtain the model parameters of the unsupervised network model and the corresponding sample data features includes:

[0015] The deep learning layer located at the first layer of the unsupervised network model is taken as the current deep learning layer;

[0016] The sample data is input into the current deep learning layer for feature extraction to obtain the layer sample data features and layer model parameters of the current deep learning layer.

[0017] The next deep learning layer is used as the current deep learning layer. The process returns to the step of inputting the sample data into the current deep learning layer for feature extraction, obtaining the layer sample data features and layer model parameters of the current deep learning layer, until the current deep learning layer is the last layer. The iteration stops when the current deep learning layer is the last layer, and the model parameters of the unsupervised network model and the sample data features corresponding to the sample data are obtained.

[0018] According to one embodiment of the fault diagnosis method for a nuclear detector of the present invention, the unsupervised network model includes multiple stacked deep learning layers; the step of adjusting the model parameters of the unsupervised network model in reverse according to the output results includes:

[0019] The sample data features and the corresponding sample fault categories are input into the backpropagation network model for supervised learning training to obtain the output results.

[0020] Based on the output results, the layer model parameters of each deep learning layer are adjusted in reverse order of the deep learning layers in the unsupervised network model to obtain the adjusted layer model parameters.

[0021] According to an embodiment of the fault diagnosis method for a nuclear detector of the present invention, the step of acquiring sample fault data and corresponding sample fault categories includes:

[0022] Acquire a set of output signals acquired by a signal acquisition device; the set of output signals includes multiple historical output signals.

[0023] Each of the aforementioned historical output signals is represented by a bifinite function;

[0024] Based on different fault categories, the function parameters of the bi-finger function are adjusted to generate fault data corresponding to each fault category; the fault data includes signal amplitude, signal time, signal interval, and signal rise time.

[0025] The fault data corresponding to each fault category is used as sample fault data, and the fault category is used as the sample fault category.

[0026] According to one embodiment of the fault diagnosis method for nuclear detectors of the present invention, the fault categories include scintillator aging fault, photomultiplier tube fault, RC deviation from normal value fault, and RC tending to infinity fault.

[0027] To achieve the above-mentioned objectives, the present invention also provides a fault diagnosis device for a nuclear detector, comprising:

[0028] The signal acquisition module is used to acquire the output signal of the nuclear detector;

[0029] The fault parameter acquisition module is used to input the output signal into a pre-trained deep belief network for fault identification and obtain the fault parameters corresponding to the output signal.

[0030] The fault category determination module is used to classify the fault parameters using a generalized likelihood ratio test classification model to obtain the fault category corresponding to the output signal.

[0031] According to an embodiment of the fault diagnosis device for a nuclear detector of the present invention, the deep belief network includes an unsupervised network model and a backpropagation network model; the device further includes a deep belief network generation module for acquiring sample data and corresponding sample categories; the sample data includes sample fault data and sample fault-free data; the sample categories include sample fault categories and sample fault-free categories; the sample data is input into the unsupervised network model for unsupervised training to obtain the model parameters of the unsupervised network model and the corresponding sample data features; the sample data features and the corresponding sample fault categories are input into the backpropagation network model for supervised learning training to obtain the output result; the model parameters of the unsupervised network model are adjusted in reverse according to the output result until the training ends.

[0032] To achieve the above-mentioned objectives, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-mentioned fault diagnosis method for nuclear detectors.

[0033] To achieve the above-mentioned objectives, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps in the above-mentioned fault diagnosis method for nuclear detectors.

[0034] This invention discloses a fault diagnosis method, apparatus, computer equipment, and storage medium for nuclear detectors. It acquires the output signal of the nuclear detector, inputs the output signal into a pre-trained deep belief network for fault identification, thereby obtaining corresponding fault parameters. Then, a generalized likelihood ratio (GLR) detection algorithm is used to determine the corresponding fault category based on the fault parameters, thus achieving fault diagnosis of the nuclear detector. By determining the fault parameters corresponding to the output signal through the deep belief network, and further using the GLR to compare the fault parameters with a threshold, the fault category can be obtained quickly and accurately for detecting faults in nuclear detectors.

[0035] The deep belief network constructed in this invention comprises an unsupervised network model and a backpropagation network model. A large amount of sample data from normal operating conditions is used as unlabeled samples to pre-train the unsupervised network model. Then, the model parameters of the unsupervised network model are optimized through backpropagation. Furthermore, various faults of nuclear detectors are simulated using virtual simulation technology to obtain a simulation dataset, which is then used as sample data to fine-tune the model parameters of the unsupervised network model. This ensures that the deep belief network accurately extracts features from multidimensional and heterogeneous nuclear signals, thereby enabling effective and accurate fault diagnosis of nuclear detectors. Furthermore, the constructed generalized likelihood ratio test classification model overcomes the variability and complexity of nuclear detector faults, exhibiting high diagnostic accuracy. It also does not require complex mathematical models to extract deep features, and overcomes problems such as overly vague diagnosis, low nonlinear efficiency, or uncontrolled convergence rate. Attached Figure Description

[0036] The following detailed description, with reference to the accompanying drawings and specific embodiments, describes the nuclear detector fault diagnosis method, apparatus, computer equipment, and storage medium of the present invention, wherein:

[0037] Figure 1 This is an application scenario diagram of the fault diagnosis method for nuclear detectors in one embodiment of the present invention;

[0038] Figure 2 This is a flowchart illustrating a fault diagnosis method for a nuclear detector, as shown in one embodiment of the present invention.

[0039] Figure 3 This is a flowchart illustrating the deep belief network in a fault diagnosis method for a nuclear detector, as shown in one embodiment of the present invention.

[0040] Figure 4 This is a flowchart illustrating a fault diagnosis method for a nuclear detector, as shown in another embodiment of the present invention.

[0041] Figure 5 This is a structural block diagram of a fault diagnosis device for a nuclear detector, as shown in one embodiment of the present invention.

[0042] Figure 6 This is an internal structural diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described in this specification are merely for explaining the invention and are not intended to limit the invention.

[0044] The fault diagnosis method for nuclear detectors provided by this invention can be applied to, for example... Figure 1 The application environment shown is illustrated. This fault diagnosis method for a nuclear detector is applied to a fault diagnosis system for a nuclear detector. The fault diagnosis system includes a nuclear detector 102, a signal acquisition device 104, a virtual simulation device 106, and a fault diagnosis device 108. The nuclear detector 102 and the signal acquisition device 104 are interconnected; the fault diagnosis device 108 is interconnected with both the virtual simulation device 106 and the signal acquisition device 104; the connection method can be wired or network-based. The nuclear detector 102 can be a NaI nuclear detector; the signal acquisition device 104 can be a high-speed acquisition card; the virtual simulation device 106 is used to encode the historical output signals to obtain sample fault categories and generate corresponding sample fault data. The signal acquisition device 104 acquires the output signal of the nuclear detector and sends it to the fault processing module 108. The fault diagnosis device 108 inputs the output signal into a pre-trained deep belief network for fault identification to obtain the corresponding fault parameters; then, a generalized likelihood ratio test classification model is used to classify the fault parameters to obtain the fault category corresponding to the output signal. The fault diagnosis device 108 can be a terminal or a server; wherein, the terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets and portable wearable devices, and the server 104 can be implemented by a standalone server or a server cluster composed of multiple servers.

[0045] Fault diagnosis systems for nuclear detectors include, for example, the KRT system. The KRT is a radiation monitoring system directly related to the operating conditions of a power plant and is an essential component of a nuclear power plant. The main targets of nuclear radiation detection include regional gamma radiation monitoring, airborne aerosol alpha and beta radiation monitoring, inert gas beta and gamma radiation monitoring, radioactive iodine gamma radiation monitoring, and liquid (water) gamma radiation monitoring. The commonly used nuclear detector is the NaI detector. Depending on usage, communication failures and detector flash alarms are more or less frequent, leading to false alarms and affecting normal production.

[0046] Therefore, after the KRT fault diagnosis system was established, it was mainly used to monitor the operating status of the nuclear detectors of concern. It can complete the detection in a timely manner when a fault occurs or is about to occur, thus achieving the purpose of rapid fault location. After the system was put into operation, it greatly expanded the fault diagnosis ideas of the KRT system, improved the system processing efficiency, and avoided catastrophic accidents. The system became the first line of defense against the spread of faults and helped maintenance personnel to buy time to find the source of the fault and eliminate it.

[0047] In one embodiment of the present invention, such as Figure 2 As shown, a fault diagnosis method for a nuclear detector is provided, which is applied to... Figure 1 Taking the fault diagnosis device 108 as an example, the following steps are included:

[0048] Step 202: Obtain the output signal of the nuclear detector.

[0049] Nuclear detectors are used in nuclear power plants to detect radiation and reflective materials, and are widely used in radiation protection and radioactivity monitoring in nuclear power plants. Nuclear detectors are characterized by high efficiency and sensitivity in detecting gamma rays and stable operation at room temperature.

[0050] The signal acquisition device acquires the output signal of the nuclear detector at a preset frequency and transmits the acquired output signal to the fault diagnosis device, which then performs fault diagnosis on the output signal.

[0051] Step 204: Input the output signal into a pre-trained deep belief network for fault identification to obtain the fault parameters corresponding to the output signal.

[0052] The deep belief network includes an unsupervised network model and a backpropagation network model. In one embodiment, the steps for generating a deep belief network include: acquiring sample data and corresponding sample categories; the sample data includes sample fault data and sample fault-free data; the sample categories include sample fault categories and sample fault-free categories; inputting the sample data into the unsupervised network model for unsupervised training to obtain the model parameters of the unsupervised network model and the corresponding sample data features; inputting the sample data features and the corresponding sample fault categories into the backpropagation network model for supervised learning training to obtain the output results; and adjusting the model parameters of the unsupervised network model in reverse according to the output results until the training is completed.

[0053] In one embodiment of the present invention, the unsupervised network model includes multiple stacked deep learning layers; inputting sample data into the unsupervised network model for unsupervised training to obtain the model parameters and corresponding sample data features of the unsupervised network model includes: taking the first deep learning layer of the unsupervised network model as the current deep learning layer; inputting sample data into the current deep learning layer for feature extraction to obtain the layer sample data features and layer model parameters of the current deep learning layer; taking the next deep learning layer as the current deep learning layer, returning to the step of inputting sample data into the current deep learning layer for feature extraction to obtain the layer sample data features and layer model parameters of the current deep learning layer, until the current deep learning layer is the last layer and stopping the iteration, thereby obtaining the model parameters of the unsupervised network model and the sample data features corresponding to the sample data.

[0054] In one embodiment of the present invention, the unsupervised network model includes multiple stacked deep learning layers; adjusting the model parameters of the unsupervised network model in reverse according to the output results includes: inputting sample data features and corresponding sample fault categories into the backpropagation network model for supervised learning training to obtain output results; and adjusting the layer model parameters of each deep learning layer according to the reverse order of the deep learning layers in the unsupervised network model based on the output results to obtain the adjusted layer model parameters.

[0055] Step 206: Use the generalized likelihood ratio test classification model to classify the fault parameters and obtain the fault category corresponding to the output signal.

[0056] The deep learning layers, which can be Restricted Boltzmann Machines (RBMs), are used to extract features from the input signal. Each deep learning layer consists of a visible layer and a hidden layer. Layer model parameters include inter-layer weights and neuron biases, i.e., the weights between the visible and hidden layers and the biases of neurons in the current layer. The backpropagation network model, which can be a backpropagation neural network, is used to fine-tune the model parameters of the unsupervised network model.

[0057] Specifically, such as Figure 3The diagram illustrates the flowchart of a deep belief network. It shows three deep learning layers, namely three RBM layers: RBM1, RBM2, and RBM3. A multi-layer network is formed by stacking RBMs, and then a contrastive divergence algorithm is used to reconstruct the input data and update the parameters, thereby constructing a pre-trained deep learning model for unsupervised deep learning. For example, the first layer, RBM1, in the unsupervised network model is used as the current deep learning layer. Sample data is input into the current deep learning layer for feature extraction, resulting in the layer sample data features and layer model parameters of the current deep learning layer. Then, the next layer, RBM2, is used as the current deep learning layer. The layer sample data features output from RBM1 are input into RBM2 for feature extraction, resulting in the corresponding layer sample data features and layer model parameters. Further, the layer sample data features output from RBM2 are input into RBM3 for feature extraction, resulting in the corresponding layer sample data features and layer model parameters. At this point, the iteration stops, and the layer model parameters of all layers of the unsupervised network model are obtained, along with the layer sample data features output from the last layer as the sample data features of the unsupervised network model. Then, the unsupervised network model is tuned using a backpropagation network model. The sample data features output by the unsupervised network model and the corresponding sample fault category output values ​​are used for supervised learning training in the backpropagation network model to obtain the output results. Then, the layer model parameters in RBM3 are adjusted sequentially according to the output results to obtain the adjusted layer model parameters in RBM3. Then, the layer model parameters in RBM2 are adjusted, and finally, the layer model parameters in RBM1 are adjusted.

[0058] In one embodiment of the present invention, a fault diagnosis method for a nuclear detector is provided, comprising the following steps:

[0059] (1) Taking the NaI detector as the research target, a detection experimental device was established. The detector output signal was acquired through a high-speed data acquisition card. The acquired signal was normalized and the fault-free output signal acquired under normal operation was used as sample data as the unsupervised learning training input of the deep learning model.

[0060] (2) Common detector faults are coded and classified, specifically including scintillator aging or damage faults, photomultiplier tube faults, and RC faults in the output analog circuit. NaI detectors utilize the principle that scintillator atoms and molecules emit fluorescence upon de-excitation, converting the light signal into electrical pulses to detect radiated particles. The main components of a scintillator detector are: a scintillator, an optical collection system (including a reflective layer), a photomultiplier tube (PMT), and voltage dividers supplying power to each stage of the PMT. These are sealed in an opaque dark box; the above devices are collectively referred to as probes. Detectors typically operate in radiation field environments, which are harsh. Long-term or excessive radiation can easily lead to various faults, which are frequent, causing flash alarms, false alarms, and affecting normal production.

[0061] Fault 1: Scintillator aging or damage fault. NaI scintillator is a transparent crystal with a T1 content of 0.1%-0.5%. Long-term operation under high-temperature irradiation will cause the scintillator detector to age, reduce the scintillator's luminous efficiency, and decrease the light output. Generally, a 50% reduction in the light output of the T1 scintillator is considered a scintillator aging or damage fault.

[0062] Fault 2: Photomultiplier tube failure. Photoelectrons emitted by the scintillator strike the photocathode of the photomultiplier tube (PMT), generating a pulse signal through the multiplier electrodes. The photomultiplier tube is a high-voltage vacuum tube; external impacts or vibrations can degrade its sealing, leading to high-voltage failure. Improper use during detector operation, such as vibration, impacts, or drops from heights, can cause photomultiplier tube failure. The following types of photomultiplier tube failures are common:

[0063] a. The ionization and excitation of residual gas inside the photomultiplier tube will increase the number and amplitude of noise when the operating voltage is high and the multiplication factor is large.

[0064] b. When the photomultiplier tube pins or sockets are damp or contaminated, leakage current will generate noise pulses. When the photomultiplier tube experiences the above-mentioned faults, the amplitude and quantity of noise in the output signal will increase several times. The noise quantity and amplitude of the NaI detector will become about 5 times the original, and due to the increased noise, the energy resolution of the pulses will become poor. RC faults in the output analog circuit: The circuit loop of the detector contains various electronic components, which can be divided into resistors and capacitors. Because the materials of these components may experience crystal cracks, aging, etc. under irradiation conditions, there are also some process damages, such as insufficient oxide layer thickness and loose solder joints. These phenomena can all cause changes in the resistance or capacitance of these components, or even open circuits. Therefore, this type of fault can be classified as an RC fault.

[0065] The detector converts the energy of incident particles into charge, forming a current pulse, which is then output as a voltage pulse through an RC circuit impedance. When the detector is used for energy measurement, its pulse amplitude can be approximated as:

[0066]

[0067] Where: Q0 is the total charge collected at the anode; τ is the decay time constant of the excited atoms (or molecules); V(t) is the output voltage pulse.

[0068] Fault 3: RC deviates from the normal value. In the circuit model, the parameters of commonly used components change by 50% as a soft fault model. Assume that the RC value in the equivalent circuit becomes 5 times the original value, and the shape of the pulse has been distorted.

[0069] Fault 4: RC approaches infinity, making it impossible to distinguish these signals as nuclear pulse signals. When RC = 0, the detector experiences a short circuit and cannot output a signal. This type of fault will not be discussed further.

[0070] (3) Construct a deep learning model with pre-trained data. The dataset of fault-free output signals collected by the data acquisition card under normal operating conditions is used as unlabeled data. Using this as input, RBM (Restricted Boltzmann Machine) is stacked to form a multi-layer network. The contrastive divergence algorithm is used to reconstruct the data and update the parameters to perform unsupervised deep learning.

[0071] like Figure 3 As shown, a DBN is composed of multiple stacked RBMs. Each RBM model consists of visible and hidden layers. Network neurons have two states: inactive and active, represented by 0 and 1 respectively. There are no self-connections between neurons within a layer, but neurons between layers are fully connected; that is, the hidden layer of RBM1 is the visible layer of RBM2, the hidden layer of RBM2 is the visible layer of RBM3, and so on. The network input is the visible layer of RBM1, and the output is the hidden layer of RBM3.

[0072] RBMs originate from undirected graphical models of Markov Random Fields (MRFs). According to the Hammersley-Clifford theorem for Markov Random Fields:

[0073]

[0074] Where C is the maximal clique in the MRF, and Z is the normalization factor, i.e., the partition function.

[0075]

[0076] Where, E(X)=∑ i∈c lnφi (x i The probability distribution is called the energy function, and its exponential representation is called the Boltzmann distribution. The energy function is a measure of the entire system. The more concentrated the probability distribution, the smaller the energy of the system. The state with the minimum energy is the most stable state.

[0077] Let m and l represent the number of neurons in the visible and hidden layers, respectively, with visible layer variable v = (v1, ..., vm) and hidden layer variable h = (h1, ..., hl). The joint probability distribution of their states (v, h) is as follows:

[0078]

[0079] because

[0080]

[0081] Therefore, the marginal probability distributions for v and h are:

[0082]

[0083]

[0084] The energy function is defined as follows:

[0085]

[0086] Where w ij represents the weight between the i-th neuron in the visible layer and the j-th neuron in the hidden layer; b = (b1, b2, ..., bm) represents the bias of the visible layer neuron; c = (c1, c2, ..., cl) represents the bias of the hidden layer neuron.

[0087] Let the energy function parameter θ = {w ij ,b j ,c i} represents the parameters of this layer of the RBM. Using log-likelihood estimation, the gradient of the function is expressed as:

[0088]

[0089]

[0090] Here This represents the RBM conditional distribution. δ(.) denotes the sigmoid function, a commonly used activation function in neural networks.

[0091]

[0092] Where <.> represents the expected value. <v j h i > data It is the expected value of the sample data. <v j h i > model It is the expectation of the reconstructed visual layer model.

[0093] A DBN is composed of multiple stacked RBMs. Each RBM has two layers: a visible layer and a hidden layer. Each layer contains many neurons, with no self-connections between neurons within the same layer and no weight relationships between neurons. The parameters of the RBM are the inter-layer weights and neuron biases. Pre-training techniques are used to initialize the weights and biases. During pre-training, the visible layers are first initialized, and then the hidden layers are calculated using the conditional distribution according to Equation 10; then, based on the hidden layers, the visible layers are calculated using the conditional distribution according to Equation 12. The result is a reconstruction of the input, and the parameters are updated by minimizing the specified error.

[0094] Δw ij =ε·( <v j h i > data - <v j h i > model (12)

[0095] Where ε represents the learning rate.

[0096] The pre-training dataset is divided into mini-batches. Each mini-batch contains a small amount of data. Parameters are updated after processing each mini-batch. DBN is pre-trained sequentially on an RBMS stack. The first hidden layer of RBM1 extracts features from the input data and passes them to RBM2, the second hidden layer extracts features from the first hidden layer, and so on. The backpropagation learning algorithm is used for training, utilizing the pre-trained weights and biases of the RBMs.

[0097] (4) Simulate common nuclear detector faults and use the fault simulation data as training data for the optimization of the unsupervised network model.

[0098] The pulse waveform output by a nuclear detector depends on the detector's characteristics, probe structure, electronics, and the type and energy of the incident particles. Statistical characteristics of nuclear detector signals:

[0099] The pulse amplitude distribution follows a normal distribution:

[0100]

[0101] Where P(A) is the probability of pulse amplitude, For the average amplitude, σ AThe amplitude standard deviation is determined by the detector's inherent energy resolution.

[0102]

[0103] The output pulse waveform can be approximated by a double exponential function:

[0104]

[0105] Where: u(t) represents the step function; A is the signal amplitude; τ2 and τ1 are the fast and slow time constants of the double exponential function, respectively, and the signal pulse decay time τ is the time constant. f and rise time τ r It is determined by τ1 and τ2. Parameters are adjusted according to different fault types to generate fault simulation data.

[0106] The detector converts the energy of incident particles into charge, forming a current pulse, which is then output as a voltage pulse through an RC circuit impedance. When the detector is used for energy measurement, its pulse amplitude can be approximated as:

[0107]

[0108] Based on the fault classification in (2), adjust the parameters and simulate the waveforms of each fault. Establish a fault information feature database, thereby constructing a labeled training dataset to assist in optimization training.

[0109] (5) A backpropagation (BP) neural network algorithm is used to assist in the tuning of the deep learning model, enabling fault identification and definition. Fault simulation data is used as labeled samples for supervised learning of the deep learning model. The parameters initialized in the pre-training stage are adjusted for the target during the training step. The parameters are updated using the gradient formula to minimize the error. In the pre-training stage, one RBM is trained at a time, while all layers of the BP neural network algorithm are trained simultaneously.

[0110] This forms a complete Deep Belief Network (DBN), such as Figure 3 As shown, it consists of multiple layers of RBM stacked together, with the output of the previous RBM serving as the input of the next RBM, and the top layer using labeled samples for auxiliary tuning.

[0111] Training deep belief networks, such as Figure 3As shown, the process includes two phases: pre-training and fine-tuning. In the pre-training phase, an unsupervised, layer-by-layer greedy training method is used to train each layer of the Restricted Block Model (RBM). The first hidden layer is trained first, and its output is used as the input to the second hidden layer, and so on, until the last hidden layer is trained. In the fine-tuning phase, the pre-trained network is fine-tuned using a supervised backpropagation (BP) algorithm to achieve global optimum.

[0112] (6) The generalized likelihood ratio test (GLRT) is used to classify faults.

[0113] The generalized likelihood ratio test (GLRT) is a composite hypothesis test that detects and classifies faults based on the magnitude of the parameters. To classify faults, three hypotheses are defined: H0: θ = θ0, H1: θ = θ1, and H2: θ = θ2. θ0, θ1, and θ2 represent no fault, gradual fault (fault 3), and abrupt fault (fault 1, fault 4), respectively.

[0114] The parameter space is defined as:

[0115]

[0116] Define Gaussian distribution:

[0117] H0={X~N(0,σ 2 I n )}

[0118] H1={X~N(θ1,σ 2 I n )}

[0119] H2={X~N(θ2,σ 2 I n )}

[0120] Perform maximum likelihood estimation on parameter θ to obtain the likelihood ratio.

[0121]

[0122] λ is called the GLRT test statistic. According to Walds' theorem, it can be reduced to...

[0123]

[0124] in Let x be the statistical mean.

[0125] In the generalized likelihood ratio test, the detection classification of a fault depends on the decision function that determines whether the hypothesis is H0 or Hi.

[0126]

[0127] Where h(α) represents the threshold of the generalized likelihood ratio.

[0128] P0(Δ(x)≥h(α))=α (20)

[0129] Where P0(B) defines the probability that event B belongs to the distribution parameter X of the null hypothesis H0. α is the probability of allowing false alarms.

[0130] In one embodiment of the present invention, the fault categories include scintillator aging fault, photomultiplier tube fault, RC deviation from normal value fault, and RC tending to infinity fault.

[0131] In a gradual fault, the fault parameters are greater than those in a normal fault but less than those in a sudden fault. This is based on the ΔW parameter of the deep belief network model. ij To assess the diagnostic accuracy of the model, data features, such as pulse amplitude values, are extracted from the deep belief network model. Faults are then classified based on the difference between the amplitude value and the average value under normal conditions, combined with the threshold of the generalized likelihood ratio. This is shown in Table 1 below.

[0132] Table 1

[0133]

[0134] In one embodiment of the present invention, such as Figure 4 The diagram illustrates a flowchart of a fault diagnosis method for nuclear detectors. First, multi-dimensional signals and samples are input into a virtual simulation device. This device encodes common fault categories of the nuclear detector, generating sample fault categories and corresponding sample fault data. The generated sample fault categories, sample fault data, and sample fault-free data are then input into a deep belief network (RBM) for fault identification. The number of deep learning layers, the number of neurons in each hidden layer, and the maximum number of iterations in the RBM are initialized. Assuming the number of layers is set to i, the RBM is trained on the input data. Unsupervised learning is performed on the current RBM layer using a contrastive divergence algorithm to obtain the output and layer model parameters. The inter-layer weights between the explicit and hidden layers, as well as the biases of the neurons in the current layer, are then saved. This process continues until the last layer is reached, at which point the output of the last RBM layer is used as the output of the unsupervised network model. Then, a backpropagation (BP) neural network is used for supervised learning to further adjust and optimize the layer model parameters of each layer in the unsupervised network model. Finally, a generalized likelihood ratio (GLR) classification model is used to classify the fault parameters output by the deep belief network, obtaining the corresponding fault category.

[0135] In the aforementioned fault diagnosis method for nuclear detectors, a deep belief network is constructed, comprising an unsupervised network model and a backpropagation network model. A large amount of sample data from normal operating conditions is used as unlabeled samples to pre-train the unsupervised network model. Then, the model parameters of the unsupervised network model are optimized through backpropagation. Furthermore, various faults of the nuclear detector are simulated using virtual simulation technology to obtain a simulation dataset, which is then used as sample data to fine-tune the model parameters of the unsupervised network model. This ensures that the deep belief network accurately extracts features from multidimensional and heterogeneous nuclear signals, thereby enabling effective and accurate fault diagnosis of the nuclear detector. Further, the constructed generalized likelihood ratio test classification model overcomes the variability and complexity of nuclear detector faults, exhibiting high diagnostic accuracy. It also does not require complex mathematical models to extract deep features, and overcomes problems such as overly vague diagnosis, low nonlinear efficiency, or uncontrolled convergence rate.

[0136] It should be understood that, although Figure 2-4 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 2-4 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0137] In one embodiment of the present invention, such as Figure 5 As shown, a fault diagnosis device 500 for a nuclear detector is provided, including: a signal acquisition module 502, a fault parameter acquisition module 504, and a fault category determination module 506, wherein:

[0138] The signal acquisition module 502 is used to acquire the output signal of the nuclear detector.

[0139] The fault parameter acquisition module 504 is used to input the output signal into a pre-trained deep belief network for fault identification and obtain the fault parameters corresponding to the output signal.

[0140] The fault category determination module 506 is used to classify fault parameters using a generalized likelihood ratio test classification model to obtain the fault category corresponding to the output signal.

[0141] In one embodiment of the present invention, the deep belief network includes an unsupervised network model and a backpropagation network model; the device further includes a deep belief network generation module for acquiring sample data and corresponding sample categories; the sample data includes sample fault data and sample fault-free data; the sample categories include sample fault categories and sample fault-free categories; the sample data is input into the unsupervised network model for unsupervised training to obtain the model parameters of the unsupervised network model and the corresponding sample data features; the sample data features and the corresponding sample fault categories are input into the backpropagation network model for supervised learning training to obtain the output results; the model parameters of the unsupervised network model are adjusted in reverse according to the output results until the training ends.

[0142] In one embodiment of the present invention, the deep belief network generation module is further configured to: take the first deep learning layer of the unsupervised network model as the current deep learning layer; input sample data into the current deep learning layer for feature extraction to obtain the layer sample data features and layer model parameters of the current deep learning layer; take the next deep learning layer as the current deep learning layer, return to the step of inputting sample data into the current deep learning layer for feature extraction to obtain the layer sample data features and layer model parameters of the current deep learning layer, and stop iterating until the current deep learning layer is the last layer, thereby obtaining the model parameters of the unsupervised network model and the sample data features corresponding to the sample data.

[0143] In one embodiment of the present invention, the deep belief network generation module is further configured to input the sample data features and the corresponding sample fault categories into the backpropagation network model for supervised learning training to obtain the output results; and adjust the layer model parameters of each deep learning layer according to the reverse order of the deep learning layers in the unsupervised network model based on the output results to obtain the adjusted layer model parameters.

[0144] In one embodiment of the present invention, the deep belief network generation module is further configured to acquire an output signal set acquired by a signal acquisition device; the output signal set includes multiple historical output signals; each historical output signal is represented by a bifinite function; the function parameters of the bifinite function are adjusted according to different fault categories to generate fault data corresponding to each fault category; the fault data includes signal amplitude, signal time, signal interval, and signal rise time; the fault data corresponding to each fault category is used as sample fault data, and the fault category is used as the sample fault category.

[0145] In one embodiment of the present invention, the fault categories include scintillator aging fault, photomultiplier tube fault, RC deviation from normal value fault, and RC tending to infinity fault.

[0146] In this embodiment, the constructed deep belief network includes an unsupervised network model and a backpropagation network model. A large amount of sample data from normal operating conditions is used as unlabeled samples to pre-train the unsupervised network model. Then, the model parameters of the unsupervised network model are optimized through backpropagation. Furthermore, various faults of the nuclear detector are simulated using virtual simulation technology to obtain a simulation dataset, which is then used as sample data to fine-tune the model parameters of the unsupervised network model. This ensures that the deep belief network accurately extracts features from multidimensional and heterogeneous nuclear signals, thereby enabling effective and accurate fault diagnosis of the nuclear detector. Further, the constructed generalized likelihood ratio test classification model overcomes the variability and complexity of nuclear detector faults, exhibiting high diagnostic accuracy. It also does not require complex mathematical models to extract deep features, and overcomes problems such as overly vague diagnosis, low nonlinear efficiency, or uncontrolled convergence rate.

[0147] Specific limitations regarding the fault diagnosis device for nuclear detectors can be found in the limitations of the fault diagnosis method for nuclear detectors mentioned above, and will not be repeated here. Each module in the aforementioned fault diagnosis device for nuclear detectors can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independently of the processor in a computer device, or stored in software in the memory of a computer device, so that the processor can call and execute the corresponding operations of each module.

[0148] In one embodiment of the present invention, a computer device is provided, which may be a fault diagnosis device, and its internal structure diagram may be as shown below. Figure 6 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores fault diagnosis data for the nuclear detector. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements a fault diagnosis method for the nuclear detector.

[0149] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0150] In one embodiment of the present invention, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described fault diagnosis method for a nuclear detector.

[0151] In one embodiment of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, wherein the computer program, when executed by a processor, implements the steps of the above-described fault diagnosis method for a nuclear detector.

[0152] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0153] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. The above embodiments only illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A fault diagnosis method for a nuclear detector, characterized in that, The fault diagnosis method includes the following steps: An experimental detection device was established, and the output signal of the nuclear detector was acquired through a high-speed data acquisition card. The acquired output signal was normalized, and the fault-free output signal acquired under normal operating conditions was used as sample data as the unsupervised learning training input of the deep learning model. The output signal is input into a pre-trained deep belief network for fault identification to obtain the fault parameters corresponding to the output signal. The deep belief network includes an unsupervised network model and a backpropagation network model. The faults of the nuclear detector are coded and classified, including faults such as aging or damage of the detector scintillator, faults of the photomultiplier tube, and RC faults in the output analog circuit. A deep learning model is constructed by pre-training data. The dataset of fault-free output signals collected by the data acquisition card under normal operating conditions is used as unlabeled data. This data is used as input, and RBM stacking is used to form a multi-layer network. The contrastive divergence algorithm is used to reconstruct the data and update the parameters to perform unsupervised deep learning. The faults of nuclear detectors are simulated, and the fault simulation data is used as the training data for the optimization of the unsupervised network model. A backpropagation (BP) neural network algorithm is used to assist in the optimization of the deep learning model, enabling the identification and definition of faults. Fault simulation data is used as labeled samples to perform supervised learning on the deep learning model. The parameters initialized in the pre-training stage are adjusted for the target in the training step, and the error is minimized by updating the parameters using the gradient formula. The fault parameters are classified using a generalized likelihood ratio test classification model to obtain the fault category corresponding to the output signal.

2. The fault diagnosis method according to claim 1, characterized in that, The steps for generating the deep belief network include: Obtain sample data and corresponding sample categories; the sample data includes faulty sample data and fault-free sample data; the sample categories include faulty sample categories; wherein fault-free sample data does not have a corresponding sample category. The sample data is input into the unsupervised network model for unsupervised training to obtain the model parameters of the unsupervised network model and the corresponding sample data features; The sample data features and the corresponding sample fault categories are input into the backpropagation network model for supervised learning training to obtain the output results. The model parameters of the unsupervised network model are adjusted in reverse based on the output results until training is complete.

3. The fault diagnosis method according to claim 2, characterized in that, The unsupervised network model includes multiple stacked deep learning layers; the step of inputting the sample data into the unsupervised network model for unsupervised training to obtain the model parameters of the unsupervised network model and the corresponding sample data features includes: The deep learning layer located at the first layer of the unsupervised network model is taken as the current deep learning layer; The sample data is input into the current deep learning layer for feature extraction to obtain the layer sample data features and layer model parameters of the current deep learning layer. The next deep learning layer is used as the current deep learning layer. The process returns to the step of inputting the sample data into the current deep learning layer for feature extraction, obtaining the layer sample data features and layer model parameters of the current deep learning layer, until the current deep learning layer is the last layer. The iteration stops when the current deep learning layer is the last layer, and the model parameters of the unsupervised network model and the sample data features corresponding to the sample data are obtained.

4. The fault diagnosis method according to claim 2, characterized in that, The unsupervised network model comprises multiple stacked deep learning layers; the step of adjusting the model parameters of the unsupervised network model in reverse based on the output results includes: The sample data features and the corresponding sample fault categories are input into the backpropagation network model for supervised learning training to obtain the output results. Based on the output results, the layer model parameters of each deep learning layer are adjusted in reverse order of the deep learning layers in the unsupervised network model to obtain the adjusted layer model parameters.

5. The fault diagnosis method according to claim 2, characterized in that, The acquisition of sample fault data and corresponding sample fault categories includes: Acquire a set of output signals acquired by a signal acquisition device; the set of output signals includes multiple historical output signals. Each of the aforementioned historical output signals is represented by a bifinite function; Based on different fault categories, the function parameters of the bi-finger function are adjusted to generate fault data corresponding to each fault category; the fault data includes signal amplitude, signal time, signal interval, and signal rise time. The fault data corresponding to each fault category is used as sample fault data, and the fault category is used as the sample fault category.

6. The fault diagnosis method according to any one of claims 1 to 5, characterized in that, The fault categories include scintillator aging fault, photomultiplier tube fault, RC deviation from normal value fault, and RC approaching infinity fault.

7. A fault diagnosis device for a nuclear detector, characterized in that, The device includes: The signal acquisition module is used to acquire the output signal of the nuclear detector; The fault parameter acquisition module is used to input the output signal into a pre-trained deep belief network for fault identification and obtain the fault parameters corresponding to the output signal. The deep belief network includes an unsupervised network model and a backpropagation network model. A deep belief network generation module is used to acquire sample data and corresponding sample categories; the sample data includes faulty sample data and fault-free sample data; the sample categories include faulty sample categories and fault-free sample categories; the sample data is input into the unsupervised network model for unsupervised training to obtain the model parameters of the unsupervised network model and the corresponding sample data features; the sample data features and the corresponding faulty sample categories are input into the backpropagation network model for supervised learning training to obtain the output results; the model parameters of the unsupervised network model are adjusted in reverse according to the output results until the training ends; The fault category determination module is used to classify the fault parameters using a generalized likelihood ratio test classification model to obtain the fault category corresponding to the output signal.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images