Nuclear power equipment weak abnormal sound detection method and computer device

Abstract

The invention belongs to the technical field of mechanical fault recognition, and particularly relates to a weak abnormal sound detection method for nuclear power equipment and a computer device, and the weak abnormal sound detection method comprises the steps: carrying out the channel fusion of a Mel-language spectrum feature matrix of extended data and a symmetric point mode feature matrix, and carrying out the training test of a Mobile-Net network model, thereby achieving the weak abnormal sound detection of the nuclear power equipment. The method and the device are used for solving the technical problems of low recognition rate of abnormal sound types of nuclear power equipment, missed judgment or misjudgment and serious consequences caused by manual sound listening detection in the prior art.

Description

Technical Field

[0001] This application belongs to the field of mechanical fault identification technology, specifically relating to a method and computer device for detecting weak abnormal noises in nuclear power equipment. Background Technology

[0002] Nuclear power equipment is primarily used in energy production, generating electricity through nuclear reactors to produce heat, providing a stable and massive power supply for large-scale electricity-consuming scenarios such as cities and industries. It is also used in non-electric applications, such as providing heat in seawater desalination, providing district heating in cold regions, and supplying high-temperature steam or hot water for industrial production processes in industries such as chemical and papermaking.

[0003] Currently, on-site condition monitoring of nuclear power equipment relies on manual listening by inspection personnel. However, the environment often involves multiple devices operating simultaneously, which can mask genuine abnormal noise sources. These unusual sounds typically indicate equipment malfunctions. Furthermore, due to the different types of malfunctions, there are various types of unusual sounds. In addition, the subjective judgment of inspection personnel is easily influenced by surrounding noise, leading to a low rate of identification of abnormal noise types from nuclear power equipment. This results in missed or incorrect diagnoses, potentially causing serious consequences. Summary of the Invention

[0004] In view of this, this application provides a method and computer device for detecting weak abnormal noises in nuclear power equipment. By using the technical means of channel fusion of the Mel spectrum feature matrix and the symmetric point pattern feature matrix of the expanded data, and the technical means of training and testing the Mobile-Net network model, this method solves the technical problem that the existing method of manual sound detection results in a low recognition rate of abnormal noise types in nuclear power equipment, leading to missed or false detections and thus serious consequences.

[0005] The first aspect of this application provides a method for detecting weak abnormal noise in nuclear power equipment, which includes the following steps 1 to 7.

[0006] Step 1: Use sound sensors to collect sound samples of normal operation of nuclear power plant equipment and abnormal sound samples of typical mechanical faults, and calibrate the sound samples of normal operation and abnormal sound samples of typical mechanical faults of various equipment.

[0007] Step 2: Use the sliding window segmentation method to adaptively obtain the sliding step size to expand the normal operation sound samples and the abnormal sound samples of typical mechanical faults of various equipment.

[0008] Step 3: Extract the Mel spectrum matrix features of sound samples from normal nuclear power plant operation and abnormal sound samples from typical mechanical faults of various equipment.

[0009] Step 4: Extract the symmetric point pattern matrix features of sound samples from normal nuclear power plant operation and abnormal sound samples from typical mechanical faults of various equipment.

[0010] Step 5: Merge the spectral features and symmetric point pattern features into a two-layer matrix feature.

[0011] Step 6: Build a Mobile-Net network using the PyTorch library. The network structure includes a feature extraction module and a feature recognition module.

[0012] Step 7: Input the fused two-layer matrix features containing normal and abnormal labels into the Mobile-Net network model to train a model for recognizing sound samples of normal nuclear power plant operation and abnormal sounds of typical mechanical faults of various equipment.

[0013] In one specific embodiment of this application, the sampling frequency in step 1 is 23759Hz, and the sampling duration is 2.4s. The calibrated data are categorized into four types: OK data, friction noise data, loosening noise data, and bearing noise data.

[0014] In one specific embodiment of this application, step 2 includes: The sliding window method is used to partition the data, and the window sliding step size is adaptively determined. The expression is as follows: in, Indicates the length of the original data. Indicates the data length of the expanded sample. This indicates the number of samples to be expanded.

[0015] For each type of abnormal sound data, the total number of consistent sample expansions is first determined, and then the sample expansion is performed.

[0016] In one specific embodiment of this application, step 3 includes steps 3-1 to 3-3.

[0017] Step 3-1: The augmented data is subjected to a short-time Fourier transform to obtain the corresponding time-frequency features, the expression of which is as follows: in, This indicates expanded data. Represents the window function. Indicates the center position of the window function. Indicates the data length. Indicates spectral line index, This represents the imaginary part of a complex exponential signal. Represents the discrete-time index of the data sequence, used to traverse the discrete-time points of the signal segment intercepted by the short-time Fourier transform.

[0018] Step 3-2 involves processing the time-frequency characteristics using a Mel filter bank to smooth the spectrum. The basic unit of the Mel filter bank is a triangular filter bank, with a center Mel frequency of... Its expression is as follows: in, For the actual frequency of the data, Indicates the first One filter, Indicates the first A Mel filter in the frequency spectrum The gain or attenuation coefficient at that point.

[0019] Step 3-3: Calculate the frequency band energy of each Mel filter output and integrate them to obtain the Mel spectrum matrix features of the augmented data, the expression of which is as follows: in, The number of Mel filters, It is a short-time Fourier transform. The center position of the window function. For spectral line indexing, Represented as a two-dimensional Mel spectrum feature matrix, the matrix size is , , The number of frames in the Melanese spectral data is expressed as: in, For data length, For the length of the window function, For the window sliding step length, This is a floor operation.

[0020] In one specific embodiment of this application, step 4 includes steps 4-1 to 4-4.

[0021] Step 4-1: Use an A-weighted filter to simulate the human ear's frequency response to sound. The formula for the A-weighted filter response is as follows: Step 4-2: Expand the data and perform A-weighted filtering, the expression of which is as follows: in, For Fast Fourier Transform, For inverse fast Fourier transform, To expand data Frequency domain data after Fourier transform, For filtered signals Time-domain data after inverse Fourier transform.

[0022] Step 4-3: Perform triangular frequency division filtering on the filtered signal, with the center filtering frequency... The expression is as follows: in, Indicates to extend the data length. This indicates the data sampling frequency.

[0023] Triangle frequency divider filter The expression is as follows: The expression for triangular frequency division filtering is as follows: in, Indicates the first Triangular filtered data, This is the time-domain data of the inverse Fourier transform.

[0024] Step 4-4: Calculate the symmetric point pattern feature matrix from the triangular filtered data. The expression is as follows: in, Indicates the first Characteristics of the filtered signal The feature matrix is ​​a symmetric point pattern matrix, and its size is the same as that of the Mel spectrogram feature matrix. same. The number of triangular filters. The radius is in polar coordinates. and These represent the angles of counter-clockwise and clockwise deflection, respectively. Indicates the rotation angle. It is the angle magnification factor and needs to satisfy... ≤ .

[0025] In one specific embodiment of this application, the feature extraction module in step 6 includes a convolutional layer, a normalization layer, a bottleneck layer, a pooling layer, and activation functions ReLU and H-Swish. The feature recognition module is a convolutional layer used for feature classification. The functions for the convolutional layer, normalization layer, bottleneck layer, and pooling layer are Conv2d(), BatchNorm2d(), Bottleneck(), and avg_pool2d() from the PyTorch library, respectively.

[0026] In one specific embodiment of this application, step 7 includes steps 7-1 to 7-4.

[0027] Step 7-1: Input the two-layer feature matrix obtained in step 5 into the Mobile-Net network model and divide the training set and test set ratios.

[0028] Step 7-2, set the loss function to cross-entropy loss. Its expression is as follows: in, Number of sample types Indicates the first The true label of the sample The model predicts that the sample belongs to the first... The probability of class label.

[0029] Step 7-3: Set the optimizer Adam, the learning rate LR, and select the batch size for each training session.

[0030] Step 7-4: Train the Mobile-Net network noise detection model using the training set, and validate the model using the test set.

[0031] A second aspect of this application provides a computer apparatus including a processor and a memory. The processor is used to execute a method for detecting weak abnormal noises in nuclear power equipment according to the first aspect of this application. The memory is used to store executable instructions of the processor.

[0032] A third aspect of this application provides a computer-readable storage medium storing executable instructions for a computer. When executed by a processor, the executable instructions implement a method for detecting weak abnormal noises in nuclear power equipment according to the first aspect of this application.

[0033] The fourth aspect of this application provides a computer program product, including a computer program / instructions, which, when executed by a processor, implements a method for detecting weak abnormal noises in nuclear power equipment according to the first aspect of this application.

[0034] The beneficial effects of the technical solution in this application are as follows: 1. By using mathematical formulas, the sliding window data segmentation method was able to achieve the technical effect of adaptively determining the sliding step size of the rectangular window, thereby achieving the technical effect of expanding and balancing the number of heterophonic data.

[0035] 2. By using the technique of channel fusion between the Melanographic feature matrix and the symmetric point pattern feature matrix of the expanded data, the technical effect of comprehensively expressing the essential features of heterophonic data was achieved.

[0036] 3. By training and testing the Mobile-Net network model, a lightweight model for detecting weak abnormal noises based on nuclear power equipment data was obtained. The model achieved a relatively high accuracy rate in identifying abnormal noise faults in nuclear power equipment. Attached Figure Description

[0037] Figure 1 The diagram shown is a flowchart illustrating a method for detecting weak abnormal noises in nuclear power equipment according to an embodiment of this application.

[0038] Figure 2 The diagram shown is a sliding window segmentation schematic of a sample expansion method provided in an embodiment of this application.

[0039] Figure 3 The diagram shown is a schematic diagram of a Mel spectrum feature matrix provided in an embodiment of this application.

[0040] Figure 4 The diagram shown is a schematic diagram of a symmetric point pattern feature matrix provided in an embodiment of this application.

[0041] Figure 5 The diagram shown is a schematic diagram of a Mobile-Net network noise detection model provided in an embodiment of this application.

[0042] Figure 6 The diagram shown is a flowchart of a method for detecting abnormal noise in nuclear power equipment according to an embodiment of this application.

[0043] Figure 7 The figure shown is a schematic diagram of the model iteration accuracy curve provided in an embodiment of this application.

[0044] Figure 8 The figure shown is a schematic diagram of the model iteration loss curve provided in an embodiment of this application.

[0045] Figure 9 The diagram shown is a classification result confusion matrix provided in an embodiment of this application.

[0046] Figure 3 and Figure 4 In the table, (1), (2), (3), and (4) represent the normal sound data, friction noise data, loosening noise data, and bearing noise data of the nuclear power equipment, respectively. Detailed Implementation

[0047] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0048] At least one embodiment of this application provides a method for detecting weak abnormal noise in nuclear power equipment. This method, based on a multi-feature fusion Mobile-Net network, is applicable to the detection of weak abnormal noise in nuclear power equipment. Figure 1 As shown, the method for detecting weak abnormal noises in nuclear power equipment based on multi-feature fusion Mobile-Net network includes the following steps 1 to 7.

[0049] Step 1: Use sound sensors to collect sound samples of normal operation of nuclear power plant equipment and abnormal sound samples of typical mechanical faults, and calibrate the sound samples of normal operation and abnormal sound samples of typical mechanical faults of various equipment.

[0050] Specifically, step 1 is the data acquisition and calibration step. Using sound sensors, sound samples from normal operation of the nuclear power equipment and abnormal sound data from typical mechanical faults are collected, and the collected data are calibrated for normal and abnormal sound types. It should be noted that normal operation sound samples can also be referred to as normal sound data or normal operation sound samples. Similarly, abnormal sound samples from typical mechanical faults can also be referred to as typical abnormal sound data or typical mechanical fault abnormal sound samples.

[0051] For example, sound sensors were used to collect sound samples of normal operation of nuclear power equipment and abnormal noise samples of typical mechanical faults. The sampling frequency was 23759 Hz, and the sampling duration was 2.4 seconds. The data was labeled into four categories: OK data, friction noise data, loosening noise data, and bearing noise data.

[0052] Step 2: Use the sliding window segmentation method to adaptively obtain the sliding step size to expand the normal operation sound samples and the abnormal sound samples of typical mechanical faults of various equipment.

[0053] Specifically, step 2 is the sample augmentation step. The original nuclear power equipment data is augmented using a sliding window segmentation method, with the following results: Figure 2 As shown.

[0054] Observe the quantity of various types of raw data, determine the sample augmentation length and the amount of augmentation, and adaptively determine the window sliding step size L using a formula. The specific operation process is as follows: in, Indicates the length of the original data. Indicates the data length of the expanded sample. Indicates the number of samples to be expanded. This indicates the window sliding step size for calculation.

[0055] Using the adaptive step size calculation formula, the adaptive step size points for the four types of sound data segmentation are 10625, 4722, 4249, and 4722, respectively.

[0056] Based on the adaptive step size segmentation of the data, the total number of samples expanded for the four classes are 150, 150, 154, and 160, respectively.

[0057] Step 3: Extract the Mel spectrum matrix features of sound samples from normal nuclear power plant operation and abnormal sound samples from typical mechanical faults of various equipment.

[0058] Step 3-1: The augmented data is subjected to a short-time Fourier transform to obtain the corresponding time-frequency feature matrix, the expression of which is as follows: in, This indicates expanded data. Represents the window function. Indicates the center position of the window function. Indicates the number of data points. Indicates spectral line index, This represents the imaginary part of a complex exponential signal. Represents the discrete-time index of the data sequence, used to traverse the discrete-time points of the signal segment intercepted by the short-time Fourier transform.

[0059] Step 3-2 involves processing the time-frequency characteristics using a Mel filter bank to smooth the spectrum. The basic unit of a Mel filter bank is a triangular filter bank, with its center frequency being the Mel frequency. Its expression is as follows: in, For the first A Mel filter, For spectral line indexes.

[0060] Step 3-3: Calculate and integrate the frequency band energy output of each Mel filter bank to obtain the Mel spectrogram features of the augmented data, expressed as follows: in, The number of Mel filters is 100. It is a short-time Fourier transform. The center position of the window function. For spectral line indexing, Represented as a Mel-based spectrogram feature matrix, the matrix size is 100×100, and the specific effect is as follows: Figure 3 As shown. The row number represents the number of Mel filters, and the column number represents the number of Mel spectrogram frames, expressed as: in, This indicates that the window function length is 512, and S indicates that the window movement step size is 278. This is a floor operation.

[0061] Step 4: Extract the symmetric point pattern matrix features of sound samples from normal nuclear power plant operation and abnormal sound samples from typical mechanical faults of various equipment.

[0062] Step 4-1: Use an A-weighted filter to simulate the human ear's frequency response to sound. The formula for the A-weighted filter response is as follows: in, This represents the actual frequency of the data.

[0063] Step 4-2: Expand the data and perform A-weighted filtering, the expression of which is as follows: in, For Fast Fourier Transform, For inverse fast Fourier transform, for Frequency domain data after Fourier transform, For filtered signals Time-domain data after inverse Fourier transform.

[0064] Step 4-3: Perform triangular frequency division filtering on the filtered signal, with the center filtering frequency... The expression is as follows: in, Indicates to extend the data length. This indicates the data sampling frequency.

[0065] Triangle frequency divider filter The expression is as follows: The expression for triangular frequency division filtering is as follows: in, For the first Each filter filters the signal. This is the time-domain data after the inverse Fourier transform.

[0066] Step 4-4: Calculate the symmetric point pattern feature matrix from the triangular filtered data. The expression is as follows: in, Indicates the first Characteristics of the filtered signal This is a feature matrix for symmetric point patterns, with a matrix size of 100×100. The specific effect is as follows: Figure 4 As shown. There are 10 triangular filters. The radius is in polar coordinates. and These represent the angles of counter-clockwise and clockwise deflection, respectively. This indicates that the rotation angle is set to 60°. It is the angle magnification factor and needs to satisfy... ≤ , For discrete-time signal points, The lag coefficient, For the whole to represent the first Based on a discrete time-domain signal point, offset The other discrete-time signal point corresponding to the first position.

[0067] For example, Set to 35°.

[0068] Step 5: Merge the Mel spectrum features and the symmetric point pattern features into a two-layer matrix feature.

[0069] Specifically, the Mel spectrum feature matrix and the symmetric point pattern feature matrix of the expanded data are fused according to the data channel direction to form a two-layer feature matrix, which is then calibrated.

[0070] For example, the 100×100 spectral feature matrix and the symmetric point pattern feature matrix are fused according to the channel dimension of the high-dimensional data to form a two-layer feature matrix. This two-layer feature matrix has a size of 2×100×100.

[0071] Step 6: Build a Mobile-Net network using the PyTorch library. The network structure includes a feature extraction module and a feature recognition module.

[0072] Specifically, within the PyTorch framework, multiple layers of convolutional layers, normalization layers, bottleneck layers, and pooling layers are stacked sequentially to construct a sound data recognition classifier with multi-label output features.

[0073] It should be noted that step 6 is the step of building the Mobile-Net network model. The Mobile-Net network model is as follows: Figure 5 As shown.

[0074] For example, the feature extraction module includes three convolutional layers for feature extraction, with kernel sizes of 3×3, 1×1, and 1×1 respectively; two normalization layers normalize the channel data extracted by the first two convolutional layers; eleven bottleneck layers are stacked into one module to reduce the number of network parameters and computational cost, while enhancing the network's feature extraction capability. Pooling layers reduce the spatial dimensionality of the data while retaining important feature information; the pooling kernel size is 7×7.

[0075] The feature recognition module consists of two convolutional layers with a kernel size of 1×1, which replace the fully connected layers for feature output.

[0076] The Mobile-Net network architecture parameters are shown in Table 1, and the operation process is as follows: Figure 6 As shown.

[0077] Table 1 Mobile-Net Network Structure Parameters The functions for the convolutional layer, normalization layer, bottleneck layer, and pooling layer are Conv2d(), BatchNorm2d(), Bottleneck(), and avg_pool2d() from the PyTorch library, respectively.

[0078] Step 7: Input the fused two-layer matrix features containing normal and abnormal labels into the Mobile-Net network model to train a model for recognizing sound samples of normal nuclear power plant operation and abnormal sounds of typical mechanical faults of various equipment.

[0079] In some embodiments, step 7 may be specifically implemented as steps 7-1 to 7-4.

[0080] Step 7-1: Input the two-layer feature matrix obtained in step 5 into the Mobile-Ne network, and divide the training set and test set into a ratio of 8:2.

[0081] Step 7-2: Set the loss function to cross-entropy loss. Its expression is as follows: in, The number of sample types is 4. Indicates the first The true label of the sample The model predicts that the sample belongs to the first... The probability of class label.

[0082] Step 7-3: Set the optimizer to Adam, the learning rate to LR, and the batch size for training to 16.

[0083] Step 7-4: Train the Mobile-Net network noise detection model using the training set, and validate the model using the test set.

[0084] In this embodiment, 75 sets of raw sound data were collected from a nuclear power plant. Using the sample augmentation method described in this invention, 614 sets of augmented data were obtained. The Mel-spectral feature matrix and symmetric point pattern feature matrix of each set were calculated and fused to obtain a dataset feature size of 614×2×100×100. The two-layer features of the augmented data were input into a Mobile-Net network, and the dataset was divided into a training set of 491 sets and a test set of 123 sets.

[0085] The model iteration accuracy curve is as follows Figure 7 As shown, the accuracy gradually increases and tends to stabilize, and the model iteration loss curve is as follows. Figure 8 As shown, the loss value decreases rapidly according to the gradient descent criterion and eventually converges. Figure 7 and Figure 8 This indicates that the model reached its optimal state after 500 iterations. The final abnormal sound recognition and classification results are as follows: Figure 9 As shown in the confusion matrix, the recognition rate for friction noise and loosening noise is 1, indicating that the model can completely identify and distinguish these two types of noise data. The recognition rates for normal sound and bearing noise are 0.933 and 0.938, respectively, showing high accuracy. The overall confusion matrix structure demonstrates that the nuclear power equipment weak noise detection model established in this invention has a good recognition effect on normal sound data and noise samples of typical mechanical faults in nuclear power equipment.

[0086] The abnormal noise recognition model designed in this application has a high recognition rate, which can assist on-site inspection personnel in detecting abnormal noises, quickly identify abnormal noise equipment on site, and reduce enterprise maintenance costs.

[0087] This application utilizes mathematical formulas to achieve the adaptive determination of the rectangular window sliding step size in the sliding window data segmentation method, thereby achieving a balanced expansion of abnormal sound data. Furthermore, by channel fusion of the Mel spectrum feature matrix and the symmetric point pattern feature matrix of the expanded data, a comprehensive and essential feature representation of the abnormal sound data is achieved. In addition, through training and testing a Mobile-Net network model, a lightweight model for weak abnormal sound detection based on nuclear power equipment data is obtained. This model achieves a relatively high accuracy rate in identifying abnormal sound faults in nuclear power equipment.

[0088] At least one embodiment of this application also provides a computer device including a processor and a memory. The processor is used to execute a method for detecting weak abnormal noises in nuclear power equipment provided in any of the above embodiments of this application. The memory is used to store executable instructions of the processor, such as application programs. The number of processors can be one or more. The application programs stored in the memory can include one or more modules, each corresponding to a set of instructions. Furthermore, the processor is configured to execute instructions to perform the above-described method for calculating the power range calibration coefficient before pressurized water reactor refueling and startup.

[0089] The computer device may also include a power supply component configured for power management of the computer device, a wired or wireless network interface configured to connect the computer device to a network, and an input / output (I / O) interface. The computer device may operate on an operating system stored in memory, such as Windows Server™, Mac OSX™, Unix™, Linux™, FreeBSD™, or similar.

[0090] At least one embodiment of this application also provides a computer-readable storage medium storing executable instructions for a computer thereon. When executed by a processor, the executable instructions implement a method for detecting weak abnormal noises in nuclear power equipment provided in any of the above embodiments of this application.

[0091] A non-transitory computer-readable storage medium, when the instructions in the storage medium are executed by the processor of the computer device, enables the computer device to perform the method for calculating the power range calibration coefficient before pressurized water reactor refueling and startup. The method for calculating the power range calibration coefficient before pressurized water reactor refueling and startup is executed by a proxy program.

[0092] Those skilled in the art will recognize that the algorithmic steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0093] At least one embodiment of this application also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements a method for detecting weak abnormal noises in nuclear power equipment provided in any of the above embodiments of this application.

[0094] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a computer program product. This computer program product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the calculation method for the power range calibration coefficient before pressurized water reactor refueling and startup according to various embodiments of this application. The aforementioned storage medium includes various media capable of storing program checksums, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0095] It should be noted that the combination of the technical features in the embodiments of this application is not limited to the combination methods described in the embodiments of this application or the combination methods described in specific embodiments. All technical features described in this application can be freely combined or combined in any way, unless they contradict each other.

[0096] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the term "comprising" only indicates that it includes the explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0097] It should be noted that the combination of the technical features in the embodiments of this application is not limited to the combination methods described in the embodiments of this application or the combination methods described in specific embodiments. All technical features described in this application can be freely combined or combined in any way, unless they contradict each other.

[0098] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications or equivalent substitutions made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for detecting weak abnormal noise in nuclear power equipment, characterized in that, include: Step 1: Use sound sensors to collect sound samples of normal operation of nuclear power plant equipment and abnormal sound samples of typical mechanical faults, and calibrate the sound samples of normal operation and abnormal sound samples of typical mechanical faults of various equipment. Step 2: Use the sliding window segmentation method to adaptively obtain the sliding step size to expand the normal operation sound samples and the abnormal sound samples of typical mechanical faults of various equipment. Step 3: Extract the Mel-spectrum matrix features of normal operation sound samples and abnormal sound samples of typical mechanical faults of various equipment; Step 4: Extract the symmetric point pattern matrix features of normal operation sound samples and abnormal noise samples of typical mechanical faults of various equipment; Step 5: Fuse the Mel spectrum features and symmetric point pattern features into a two-layer matrix feature; Step 6: Build a Mobile-Net network using the PyTorch library. The network structure includes a feature extraction module and a feature recognition module. Step 7: Input the fused two-layer matrix features containing normal and abnormal labels into the Mobile-Net network model to train the nuclear power plant normal and typical abnormal sound data recognition model.

2. The method for detecting weak abnormal noise in nuclear power equipment according to claim 1, characterized in that, In step 1, the sampling frequency is 23759Hz, the sampling duration is 2.4s, and the calibrated data are categorized into four types: OK data, friction noise data, loosening noise data, and bearing noise data.

3. The method for detecting weak abnormal noise in nuclear power equipment according to claim 1, characterized in that, Step 2 includes: The sliding window method is used to segment data, and the window sliding step size L is adaptively determined. Its expression is as follows: in, Indicates the length of the original data. Indicates the data length of the expanded sample. Indicates the number of samples to be expanded. For each type of abnormal sound data, the total number of consistent sample expansions is first determined, and then the sample expansion is performed.

4. The method for detecting weak abnormal noise in nuclear power equipment according to claim 1, characterized in that, Step 3 includes: Step 3-1: The augmented data is subjected to a short-time Fourier transform to obtain the corresponding time-frequency features, the expression of which is as follows: Where x(n) represents the augmented data, ω represents the window function, τ represents the center position of the window function, N represents the data length, and k represents the spectral line index. This represents the imaginary part of a complex exponential signal. Represents the discrete-time index of the data sequence, used to traverse the discrete-time points of the signal segment intercepted by the short-time Fourier transform; Step 3-2: Process the time-frequency characteristics using a Mel filter bank to smooth the spectrum. The basic unit of the Mel filter bank is a triangular filter bank with a center Mel frequency of f(mel), and its expression is as follows: Where f is the actual frequency of the data, m represents the m-th Mel filter, and Hm(k) represents the gain or attenuation coefficient of the m-th Mel filter at frequency line k. Step 3-3: Calculate the frequency band energy of each Mel filter output and integrate them to obtain the Mel spectrum matrix features of the augmented data, the expression of which is as follows: Where M is the number of Mel filters, It is a short-time Fourier transform. The center position of the window function. For spectral line index, S mel Represented as a two-dimensional Mel spectrogram feature matrix, the matrix size is a×b, where a=M and b is the number of Mel spectrogram frames, and its expression is: Where N is the data length, W is the window function length, and S is the window sliding step size. This is a floor operation.

5. The method for detecting weak abnormal noise in nuclear power equipment according to claim 1, characterized in that, Step 4 includes: Step 4-1: Use an A-weighted filter to simulate the human ear's frequency response to sound. The A-weighted filter response... The formula is as follows: ， Where f is the actual frequency of the data; Step 4-2: Expand the data and perform A-weighted filtering, the expression of which is as follows: in, For Fast Fourier Transform, For the inverse fast Fourier transform, X(f) is the frequency domain data of the extended data x(n) after Fourier transform, and y(n) is the time domain data of the filtered signal Y(f) after inverse Fourier transform. Step 4-3: Perform triangular frequency division filtering on the filtered signal. The expression for the center filtering frequency f(m) is as follows: ， Where N represents the extended data length, f s Indicates the data sampling frequency. Triangle frequency divider filter The expression is as follows: The expression for triangular frequency division filtering is as follows: Among them, Z m (f) represents the m-th triangular filter data, z m (n) represents the time-domain data of the inverse Fourier transform; in step 4-4, the symmetric point pattern feature matrix is ​​calculated from the triangular filtered data, and its expression is as follows: Among them, SDP m S represents the characteristic of the m-th filtered signal. SDP This is the symmetric point pattern feature matrix, with the same size (a×b) as the Mel spectrum feature matrix. M' represents the number of triangular filters, r is the polar radius, and θ and φ represent the counterclockwise and clockwise deflection angles, respectively. l Let ξ represent the rotation angle, and let ξ be the angle magnification factor, which must satisfy ξ≤θ. For discrete-time signal points, The lag coefficient, For the whole to represent the first Based on a discrete time-domain signal point, offset The other discrete-time signal point corresponding to the first position.

6. The method for detecting weak abnormal noise in nuclear power equipment according to claim 1, characterized in that, Step 6's feature extraction module includes convolutional layers, normalization layers, bottleneck layers, pooling layers, and activation functions ReLU and H-Swish. The feature recognition module is a convolutional layer used for feature classification. The functions for the convolutional layers, normalization layers, bottleneck layers, and pooling layers are Conv2d(), BatchNorm2d(), Bottleneck(), and avg_pool2d() from the PyTorch library, respectively.

7. The method for detecting weak abnormal noise in nuclear power equipment according to claim 1, characterized in that, Step 7 includes: Step 7-1: Input the two-layer feature matrix obtained in step 5 into the Mobile-Net network model and divide the training set and test set into proportions. Step 7-2, set the loss function to cross-entropy loss, and the expression for the loss function L is as follows: Where q is the number of sample types, y i p represents the true label of the i-th class sample. i This indicates the probability that the model predicts the sample belongs to the i-th label class; Step 7-3: Set the optimizer Adam, the learning rate LR, and select the batch size for each training session; Step 7-4: Train the Mobile-Net network noise detection model using the training set, and validate the model using the test set.

8. A computer device, characterized in that, include: A processor for executing a method for detecting weak abnormal noises in nuclear power equipment according to any one of claims 1 to 7; as well as Memory for storing the executable instructions of the processor.

9. A computer-readable storage medium having executable instructions stored thereon, characterized in that, When the executable instructions are executed by the processor, they implement the method for detecting weak abnormal noises in nuclear power equipment as described in any one of claims 1 to 7.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the method for detecting weak abnormal noises in nuclear power equipment as described in any one of claims 1 to 7.

Images