Green ammonia reactor fault mode identification method and device

Abstract

The invention relates to the technical field of data processing and data recognition, in particular to a green ammonia reactor fault mode recognition method and device. The method comprises the following steps: acquiring target characteristic time sequence data of a target green ammonia reactor in a target time period; wherein the target characteristic time sequence data comprises target vibration time sequence data and target temperature time sequence data; feature enhancement processing and feature fusion processing are sequentially carried out on the target vibration time sequence data and the target temperature time sequence data to obtain a target dimension reduction temperature vibration fusion feature vector; according to the target dimension reduction temperature vibration fusion feature vector, determining a fault mode category corresponding to the target feature time sequence data through a pre-trained fault mode recognition model; the technical problem that the fault mode category of the green ammonia reactor is difficult to accurately judge in the prior art can be solved.

Description

Technical Field

[0001] The present application relates to the technical field of data processing and data identification, and in particular to a method and device for identifying a fault pattern of a green ammonia reactor. Background Art

[0002] In the field of renewable energy, the green hydrogen industry chain is a series of industrial links formed around the development, storage, transportation, application and recycling of green hydrogen produced through renewable energy (such as wind power, photovoltaics, hydropower, etc.); among them, the green ammonia reactor is the core equipment in the green hydrogen industry chain, which is used to synthesize nitrogen (N2) and hydrogen (H2) into ammonia (NH3) through a chemical reaction driven by renewable energy, thereby realizing "zero-carbon" or "low-carbon" ammonia production.

[0003] In actual operation, green ammonia reactors operate in high-temperature, high-pressure and complex chemical reaction environments for a long time in industrial processes such as hydrogen production and ammonia synthesis. Their internal heat distribution, mechanical vibration and material flow state are highly coupled. During their service, green ammonia reactors are prone to various failure modes such as seal leakage, interface leakage, local or global blockage of the catalyst bed, bearing wear and rotor imbalance.

[0004] In actual applications, although there are many failure modes of green ammonia reactors, all of the above failure modes show small local gradient changes in the temperature field or subtle changes in narrow-band energy in the vibration signal in the early stage. The signal amplitude is low and has the limitation of being easily masked by environmental noise and process fluctuations. Therefore, it is difficult to accurately judge the failure mode category of green ammonia reactors with existing technology. Summary of the Invention

[0005] In view of this, the purpose of the present application is to provide a method and device for identifying a fault mode of a green ammonia reactor, so as to solve the technical problem in the prior art that it is difficult to accurately judge the fault mode category of a green ammonia reactor.

[0006] In a first aspect, the present application provides a method for identifying a fault mode of a green ammonia reactor, the method comprising: Obtain target characteristic time series data of a target green ammonia reactor within a target time period; wherein the target characteristic time series data includes: target vibration time series data and target temperature time series data; For the target vibration time series data and the target temperature time series data, feature enhancement processing and feature fusion processing are sequentially performed to obtain a target dimension-reduced temperature-vibration fusion feature vector; According to the target dimension-reduced temperature-vibration fusion feature vector, a fault mode category corresponding to the target feature time series data is determined through a pre-trained fault mode recognition model.

[0007] In a second aspect, the present application provides a green ammonia reactor fault pattern recognition device, the device comprising: a data acquisition module, a data processing module and a prediction module; The data acquisition module is used to obtain target characteristic time series data of the target green ammonia reactor within a target time period; wherein the target characteristic time series data includes: target vibration time series data and target temperature time series data; The data processing module is used to perform feature enhancement processing and feature fusion processing on the target vibration time series data and the target temperature time series data in sequence to obtain a target dimensionality-reduced temperature-vibration fusion feature vector; The prediction module is used to determine the fault mode category corresponding to the target feature time series data through a pre-trained fault mode recognition model based on the target dimensionality reduction temperature-vibration fusion feature vector.

[0008] Beneficial effects: After collecting two types of time-series data, vibration and temperature, this application utilizes the sensitivity of vibration time-series data to mechanical failures and combines it with the response characteristics of temperature time-series data to thermodynamic anomalies to form a complementary monitoring system. This multi-source data strategy significantly improves the visibility of fault characteristics and avoids misjudgments caused by changes in operating conditions caused by a single sensor. Therefore, it can solve the problem that the early signals of green ammonia reactor failures are weak and susceptible to interference. For example, temperature gradient changes may be masked by process thermal fluctuations, and narrow-band vibration energy fluctuations are easily mixed with environmental noise. This application extracts key features from vibration time series data and temperature time series data through feature enhancement processing, and then fuses them into low-dimensional feature vectors through a dimensionality reduction algorithm. This process not only retains fault-related features but also eliminates noise interference, allowing early weak anomalies to form distinguishable patterns in the feature space, laying the foundation for subsequent accurate classification. This can solve the problem that the original time series data contains a large amount of redundant information, making it difficult to extract effective fault features through direct analysis. The fault pattern recognition model pre-trained in this application can automatically match feature time series data with fault pattern categories, and can maintain high recognition accuracy even in the face of variables such as process fluctuations or equipment aging. This classification method completely solves the technical problem of low amplitude and easy confusion of early fault signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments of the present application. The following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0010] Figure 1A flow chart of a method for identifying a green ammonia reactor failure mode provided in an embodiment of the present application; Figure 2 Another flow chart of the green ammonia reactor fault mode identification method provided in an embodiment of the present application; Figure 3 A diagram showing the impact of the group normalization strategy on feature quality provided in an embodiment of the present application; Figure 4 A comparison chart of the feature enhancement effect of vibration time series data provided by an embodiment of the present application; FIG5( a ) is a frequency band analysis diagram of sample vibration time series data provided in an embodiment of the present application; FIG5( b ) is a frequency band analysis diagram of traditional wavelet packet decomposition provided by an embodiment of the present application; FIG5( c ) is a frequency band analysis diagram of the vibration feature enhancement method provided in an embodiment of the present application; Figure 6 Comparative diagrams showing the detection effects of the temperature-enhanced feature extraction method provided in an embodiment of the present application on local blockage; wherein (a) is a schematic diagram showing the performance of the sliding average method provided in an embodiment of the present application when detecting a low-temperature zone caused by local blockage; (b) is a schematic diagram showing the performance of the global polynomial fitting method provided in an embodiment of the present application when detecting a low-temperature zone caused by local blockage; (c) is a schematic diagram showing the performance of the gradient weighted method provided in an embodiment of the present application when detecting a low-temperature zone caused by local blockage; Figure 7 A comparison chart of different fault mode identification methods provided in the embodiments of the present application; Figure 8 This is a structural diagram of the green ammonia reactor fault mode identification device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0011] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions of this application will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0012] First, this application provides a method for identifying fault patterns in a green ammonia reactor, such as Figure 1 As shown, Figure 1 A flow chart of a method for identifying a green ammonia reactor failure mode provided in an embodiment of the present application, the method includes: S210 to S230, details of which are as follows: S210: Acquire target characteristic time series data of a target green ammonia reactor within a target time period.

[0013] The target characteristic time series data includes target vibration time series data and target temperature time series data.

[0014] Specifically, in the embodiment of the present application, the "target green ammonia reactor" is a green ammonia reactor with a "fault mode identification category" requirement; the "target feature time series data" is the basic data required to identify the fault mode category of the target green ammonia reactor.

[0015] In actual operation, high-precision temperature sensor arrays and vibration sensor arrays are deployed at key locations in the target green ammonia reactor. The temperature sensors are evenly distributed on the pipe wall of the target green ammonia reactor in the form of a linear array, with a spatial density of one monitoring point every 5 cm. The vibration sensor is installed at the connection between the bearing support point and the pipeline in the target green ammonia reactor.

[0016] In actual operation, the temperature and vibration signals are acquired synchronously in real time through the industrial data acquisition system, with a sampling frequency of 1024 Hz, and the target feature time series data is obtained by continuous acquisition for 4 seconds.

[0017] Each target feature time series data contains 4096 dimensions of original data. The first 2048 dimensions of the target feature time series data are target temperature time series data, that is, the temperature values ​​collected at 2048 sampling moments, which are arranged in a time sequence; the last 2048 dimensions of the target feature time series data are target vibration time series data. In actual operation, the target vibration time series data is the composite amplitude of three-dimensional vibration acceleration; wherein, in three-dimensional space, the vibration acceleration of an object can be decomposed into three mutually perpendicular components, and the composite amplitude is the modulus of the vector sum of the three components, reflecting the total intensity of the vibration. Therefore, the target vibration time series data is the composite amplitude of the acceleration collected at 2048 sampling moments, which is arranged in a time sequence.

[0018] It should be emphasized that the temperature value and the acceleration composite amplitude are collected synchronously, that is, the 2048 sampling moments corresponding to the target vibration time series data and the target temperature time series data should be the same, that is, at each sampling moment, a temperature value and an acceleration composite amplitude are collected at the same time, and then combined into the target feature time series data.

[0019] In an embodiment of the present application, after the target feature time series data is obtained, the target feature time series data is preprocessed by feature enhancement processing, and then the data obtained after preprocessing is input into a pre-trained fault pattern recognition model. The data output by the fault pattern recognition model is the fault mode category corresponding to the target feature time series data, and the fault mode category indicates the category to which the fault mode of the target green ammonia reactor belongs.

[0020] In one implementation, before S210, as Figure 2 As shown, Figure 2 Another flow chart of the green ammonia reactor fault mode identification method provided in an embodiment of the present application, the method further includes: S110 to S160, the details of which are as follows: S110: Acquire multiple sample characteristic time series data of the sample green ammonia reactor within a sample time period, wherein the sample characteristic time series data includes: sample vibration time series data and sample temperature time series data.

[0021] Specifically, in the embodiment of the present application, before applying the fault mode recognition model, it is necessary to perform iterative training on the fault mode recognition model.

[0022] In actual operation, the temperature and vibration signals are acquired synchronously in real time through the industrial data acquisition system with a sampling frequency of 1024 Hz. A sample feature time series data is obtained by continuous sampling for 4 seconds. The data is collected continuously for 3 months, and then 8,000 sample feature time series data are extracted from the sample feature time series data collected in 3 months. Among them, the number of positive samples and negative samples is 4,000.

[0023] Each sample feature time series data contains 4096 dimensions of original data. The first 2048 dimensions of the sample feature time series data are sample temperature time series data, that is, the temperature values ​​collected at 2048 sampling moments, a sequence composed of time series; the last 2048 dimensions of the sample feature time series data are sample vibration time series data. In actual operation, the sample vibration time series data is the composite amplitude of three-dimensional vibration acceleration; wherein, in three-dimensional space, the vibration acceleration of an object can be decomposed into three mutually perpendicular components, and the composite amplitude is the modulus of the vector sum of the three components, reflecting the total intensity of the vibration. Therefore, the sample vibration time series data is the composite amplitude of the acceleration collected at 2048 sampling moments, a sequence composed of time series.

[0024] In an embodiment of the present application, after collecting multiple sample feature time series data that meet the model training requirements, each sample feature time series data is labeled according to a preset label; wherein, the labeling of the fault mode category is determined based on the green ammonia reactor operation log and the expert diagnosis results, and the label indicates the fault mode category.

[0025] In actual operation, when the sample feature time series data indicates that the green ammonia reactor is in normal operation, it is labeled as the "normal" category; when the sample feature time series data indicates that a gas leak is detected, it is labeled as "seal leakage" or "interface leakage" according to the leak location (sealing ring / flange interface); when the sample feature time series data indicates that an abnormality in the catalyst bed is detected, it is labeled as "local blockage" or "global blockage" according to the pressure difference change; when the sample feature time series data indicates that an abnormality in a mechanical component is detected, it is labeled as "bearing wear" or "rotor imbalance" according to the spectrum characteristics, forming a total of 7 types of fault labels.

[0026] S120: According to the noise characteristics corresponding to the vibration time series data and the temperature time series data of the green ammonia reactor, the sample temperature time series data and the sample vibration time series data in each sample feature time series data are respectively normalized corresponding to the noise characteristics to obtain a sample normalized temperature sub-vector and a sample normalized vibration sub-vector.

[0027] Specifically, in actual operation, the vibration time series data and temperature time series data of the green ammonia reactor have different noise characteristics. The noise characteristics of the vibration time series data are high-frequency noise and non-stationary fluctuations, and the noise characteristics of the temperature time series data are smooth and low noise. There are significant differences in the physical properties and statistical distribution of the two. Therefore, if the conventional global normalization method is used to normalize the vibration time series data and the temperature time series data, the subtle changes in the temperature time series data will be submerged by the high-amplitude noise of the vibration time series data, destroying the fault-sensitive characteristics such as the local temperature gradient corresponding to fault modes such as leakage or blockage, thereby reducing the recognition accuracy of the fault mode category.

[0028] In order to solve the above technical problems, the embodiment of the present application adopts a group normalization strategy, and independently normalizes the heterogeneous characteristics of the temperature time series data and the vibration time series data according to the noise characteristics corresponding to the vibration time series data and the temperature time series data of the green ammonia reactor.

[0029] In one implementation, step S120 includes steps (1) to (3), the details of which are as follows: Step (1): According to the sample temperature time series data and the sample vibration time series data in each sample feature time series data, determine the sample vibration time series data sub-vector and the sample temperature time series data sub-vector corresponding to each sample feature time series data.

[0030] Specifically, in an embodiment of the present application, a sample temperature time series data sub-vector is obtained based on the first 2048-dimensional elements of the sample feature time series data, and a sample vibration time series data sub-vector is obtained based on the elements from the 2049th dimension to the 4096th dimension, thereby solving the problem of physical scale mismatch.

[0031] In actual operation, define the sample temperature time series data subvector , the dimension of the sample temperature time series data sub-vector , sample vibration time series data sub-vector , the dimension of the sample vibration time series data sub-vector .

[0032] Where, Represents the sample characteristic time series data of the sample green ammonia reactor, with a dimension of 4096, consisting of the sample temperature time series data subvector and the sample vibration time series data subvector Spliced ​​together; Represents sample feature time series data Before Dimensional elements; Represents sample feature time series data No. Dimension to the first Dimensional elements.

[0033] Step (2): According to the noise characteristics corresponding to the vibration time series data of the green ammonia reactor, the sample vibration time series data subvector is normalized to obtain a sample normalized vibration subvector.

[0034] Specifically, a sample normalized vibration time series data subvector is obtained based on the mean and standard deviation of the sample vibration time series data, thereby suppressing high-frequency noise and standardizing the distribution. The formula for normalizing the sample vibration time series data is as follows: ; ; ; Where, Represents the sample normalized vibration time series data subvector; Represents the average value of multiple acceleration composite amplitudes included in the sample vibration time series data; Indicates the sample vibration time series data included The first of the acceleration composite amplitudes The composite amplitude of acceleration; ; Indicates the standard deviation of multiple acceleration composite amplitudes included in the sample vibration time series data; Indicates a very small constant, used to prevent division by zero errors. In actual operation, The value is set to .

[0035] Step (3): According to the noise characteristics corresponding to the temperature time series data of the green ammonia reactor, the sample temperature time series data subvector is normalized to obtain a sample normalized temperature subvector.

[0036] Specifically, the sample normalized temperature time series data subvector is obtained based on the 10th percentile and 90th percentile of the sample temperature time series data, thereby eliminating the influence of outliers and retaining the smoothness feature. The formula for normalizing the sample temperature time series data is as follows: ; Where, Represents the sample normalized temperature time series data subvector; Indicates the 10th percentile of the sample temperature time series data; Indicates the 90th percentile of the sample temperature time series data.

[0037] In the embodiment of the present application, when the sample normalized vibration time series data subvector is obtained and the sample normalized temperature time series data subvector After that, you can and Connect to the sample recombinant vector by vector splicing , with a dimension of 4096.

[0038] It should be noted that the embodiment of the present application abandons the conventional global normalization homogenization processing and solves the core contradiction of multimodal data fusion through the strategy of independent normalization + splicing and reorganization of heterogeneous features. The temperature time series data adopts quantile-based normalization, which is essentially to isolate abnormal points using robust statistics. The vibration time series data adopts difference normalization with a smoothing factor. The two types of features are independently processed and spliced, which produces a cross-modal feature complementary effect in fault mode recognition. When the reactor is partially blocked, the vibration energy distribution in the area where the temperature gradient rises will show a specific frequency band attenuation. The reorganized sample reorganization vector Make the association pattern linearly separable in the feature space.

[0039] For example, a comparative analysis of the effects of the group normalization strategy is conducted to verify the effect of the group normalization strategy on improving the quality of fault features. The group normalization strategy is a strategy for normalizing the sample vibration time series data and the sample temperature time series data respectively, such as Figure 3 As shown, Figure 3 This is a diagram showing the impact of the group normalization strategy on feature quality provided in the embodiment of this application. Figure 3 The horizontal axis is the label corresponding to the 7 fault identification categories. Figure 3The vertical axis is the feature discrimination (F1 score), which is a comprehensive indicator used in machine learning and statistics to evaluate the performance of classification models.

[0040] Figure 3 The feature discrimination of four methods, namely global normalization, temperature group normalization, vibration group normalization and the present technology (combination), on seven types of fault modes was compared. The "present technology (combination)" refers to the group normalization strategy provided in the embodiment of the present application. The experimental results show that global normalization (the first column in each column group) performs the worst in all fault mode categories, with significantly low feature discrimination, because the high-amplitude noise of the vibration time series data drowns out the subtle changes in the temperature time series data, resulting in the destruction of fault-sensitive features. The group normalization strategy provided in the embodiment of the present application (the last column in each column group) achieves the highest feature discrimination for all types of faults, especially in "interface leakage" and "bearing wear". This shows that quantile normalization is used to isolate abnormal points in temperature time series data, and difference normalization is used to suppress high-frequency noise in vibration time series data. The two are processed independently and then spliced ​​and recombined, making the cross-modal correlation features of temperature gradient changes and vibration energy distribution linearly separable in the feature space.

[0041] S130: Perform feature enhancement processing on the sample normalized vibration sub-vector corresponding to each sample feature time series data to obtain a sample vibration enhancement feature.

[0042] Specifically, in an embodiment of the present application, the sample feature time series data contains high-frequency components and transient impact signals that are strongly correlated with mechanical faults, but are easily affected by environmental noise. Conventional technologies use raw time domain data or simple fast Fourier transforms for frequency domain transformation, and are unable to adaptively extract key frequency bands related to faults, resulting in insufficient sensitivity to subtle fault modes such as early cracks or bearing wear, and key fault features are easily submerged by noise.

[0043] In order to solve the above technical problems, the embodiment of the present application adopts a fusion method of adaptive bandpass filtering and wavelet packet decomposition to enhance the sample feature time series data.

[0044] In one implementation, the sample vibration time series data includes a plurality of sample acceleration composite amplitudes corresponding to a plurality of sampling moments; S130 includes: steps (4) to (7), the details of which are as follows: Step (4): Based on the amplitudes of the sample vibration time series data at the sample acceleration composite amplitudes with different values ​​in each sample feature time series data, clustering processing is performed on the sample vibration time series data to obtain the number of sample frequency bands corresponding to the sub-frequency bands of each sample normalized vibration sub-vector.

[0045] Specifically, in the embodiment of the present application, it is necessary to divide the sample feature time series data into multiple sub-band sets based on the number of sample frequency bands and the sample acceleration synthesis amplitude, so as to further determine the energy-significant sub-bands; wherein, the formula for determining the number of sample frequency bands is as follows: ; Where, Represents the number of sample bands, that is, the number of sample sub-bands is In actual operation, the number of sample bands Determined through cluster optimization, in practical applications, The value of can be 8; Indicates the composite amplitude of multiple sample accelerations included in the sample feature time series data. The first frequency point The composite amplitude of sample acceleration at the frequency; Indicates the number of frequency points corresponding to the sample feature time series data, that is, the number of frequency points is indivual; ; express The sample sub-band A frequency set of sub-bands; express The sample sub-band The centroid frequency of each sub-band; represents the L2 norm; The formula for minimizing the objective function "the number of sample bands" is expressed as value.

[0046] Step (5): According to the number of sample frequency bands, perform wavelet packet decomposition on each sample normalized oscillator vector to obtain the sample wavelet packet frequency band coefficient vector of each sub-band.

[0047] The sample wavelet packet frequency band coefficient vector includes: a sample wavelet packet coefficient vector in one-to-one correspondence with a plurality of vibration frequencies included in the sub-frequency band.

[0048] Specifically, in the embodiment of the present application, the sample normalized vibration time series data sub-vector Perform wavelet packet decomposition, specifically using Daubechies wavelet basis, and define The sample wavelet packet band coefficient vector of the sub-band is , based on the wavelet packet decomposition function, the wavelet packet coefficient vector of each sub-band is obtained, thereby extracting the characteristics of each sub-band.

[0049] Step (6): Determine the sample adaptive weight corresponding to each sub-band according to the plurality of sample wavelet packet coefficient vectors corresponding to each sub-band.

[0050] Specifically, in the embodiment of the present application, the total energy of each sub-band is calculated, and then the adaptive weight is obtained based on the elements of the sub-band wavelet packet coefficient vector, thereby generating the weight for strengthening the fault-sensitive frequency band; wherein, the formula for determining the sample adaptive weight is as follows: ; ; Where, Indicates the Sample adaptive weights for sub-bands; Indicates the The total energy of the sub-bands; Indicates the The total energy of the sub-band; The sub-band is sub-bands with different sub-bands; Indicates the The sample wavelet packet band coefficient vector of sub-bands , including The first of the sample wavelet packet coefficient vectors wavelet packet coefficients; .

[0051] Step (7): Determine the sample vibration enhancement feature according to the sample wavelet packet frequency band coefficient vector and sample adaptive weight corresponding to each sub-band.

[0052] Specifically, in the embodiment of the present application, the sample vibration enhancement feature after dimensionality reduction is obtained based on the sample wavelet packet frequency band coefficient vector and the first dimensionality reduction projection matrix; wherein, the sample vibration enhancement feature is defined as ; Where, Represents the first dimensionality reduction projection matrix, the first dimensionality reduction projection matrix The model parameters trained during the training of the fault pattern recognition model are generated through random Gaussian distribution initialization. Elements are sampled from a Gaussian distribution with a mean of 0 and a standard deviation of 0.01. Backpropagation optimization is used during training to achieve dimensionality reduction to reduce feature redundancy and computational complexity while retaining the main fault information. Indicates the The sample wavelet packet band coefficient vector of sub-bands, represents the sample wavelet packet band coefficient vector of the first sub-band, represents the sample wavelet packet band coefficient vector of the second sub-band, Indicates the The sample wavelet packet band coefficient vector of sub-bands; Indicates the The sample adaptive weights of the sub-bands, Indicates the Sample adaptive weights for subbands, represents the sample adaptive weight of the second sub-band, Indicates the Sample adaptive weights for sub-bands.

[0053] It should be noted that conventional wavelet packet decomposition treats all sub-bands equally, but the bearing crack fault of the green ammonia reactor will produce a narrow-band resonance concentrated in 2~4kHz. The term gives higher weight to the fault-sensitive frequency band by weighting the energy proportion. When the Daubechies wavelet basis is used, the energy concentration of the fault frequency band is superlinearly related to the weight, that is, The nonlinear amplification effect enhances the characteristic response intensity of early cracks, which cannot be mined by conventional equal-weight processing and cannot be inspired by conventional technology. More importantly, the sample adaptive weight Compute the projection matrix with the first dimensionality reduction Forming cascade optimization, after the high-frequency band weight is strengthened, the first dimension reduction projection matrix The sparse projection direction of the fault frequency band is automatically learned during training, which can achieve self-suppression of the noise band and still extract weak periodic impact components under strong background noise.

[0054] For example, a comparative analysis of time-frequency feature enhancement is now conducted to demonstrate the differences between conventional fast Fourier transform in vibration signal processing and the vibration feature enhancement method provided by the embodiment of the present application. Figure 4 As shown, Figure 4 This is a comparison chart of the feature enhancement effect of vibration time series data provided by the embodiment of the present application. Figure 4 The horizontal axis is frequency (Hz), Figure 4 The vertical axis is the normalized energy (dimensionless).

[0055] Figure 4 The conventional FFT (Fast Fourier Transform) method and the time-frequency enhancement of this technology (the vibration feature enhancement method provided in the embodiment of this application) are compared. Figure 4The spectrum energy distribution of the upper dotted line) is uniform and fluctuates violently. There is no significant characteristic peak in the key fault frequency band (170-190 Hz bearing wear sensitive band, 400-440 Hz rotor imbalance sensitive band), indicating that the high-frequency noise has submerged the fault characteristics; the vibration feature enhancement method provided in the embodiment of the present application ( Figure 4 The solid line at the bottom shows sharp energy peaks in the fault-sensitive frequency band, such as higher amplitudes at 180 Hz and 420 Hz. This indicates that after wavelet packet decomposition, the fault-sensitive frequency band is given a higher weight through energy proportion. Combined with the noise suppression function of the dimensionality reduction projection matrix, the narrowband resonance of the early bearing crack is nonlinearly amplified.

[0056] For another example, a three-dimensional spectrum waterfall diagram is used to perform frequency band analysis for vibration feature enhancement. The processing effects of different vibration feature extraction methods in the time-frequency domain are intuitively presented based on the three-dimensional spectrum waterfall diagram, as shown in Figures 5(a), 5(b) and 5(c). Figure 5(a) is a frequency band analysis diagram of the sample vibration time series data provided in an embodiment of the present application, Figure 5(b) is a frequency band analysis diagram of the traditional wavelet packet decomposition provided in an embodiment of the present application, and Figure 5(c) is a frequency band analysis diagram of the vibration feature enhancement method provided in an embodiment of the present application. In the figure, the horizontal axis represents frequency (in Hertz), the vertical axis represents time point, and the vertical axis represents energy intensity (dimensionless relative value).

[0057] Figures 5(a), 5(b), and 5(c) compare the spectral characteristics of sample vibration time series data, traditional wavelet packet decomposition, and the feature enhancement method provided by the embodiment of the present application when processing the same section of bearing wear fault data. Analysis of Figures 5(a), 5(b), and 5(c) reveals significant differences. In Figure 5(a), the fault characteristic frequency band of 150-250 Hz is drowned out by background noise. Figure 5(b) shows a relatively uniform energy distribution, but the fault-sensitive frequency band is not enhanced. Figure 5(c) demonstrates significant feature selectivity, with energy significantly enhanced in the fault-sensitive frequency band of 150-250 Hz (the characteristic frequency band of bearing cracks), forming a prominent "energy ridge." Meanwhile, energy in the noise band of 400-500 Hz is effectively suppressed. These time-frequency characteristics demonstrate that the vibration feature enhancement method provided by the embodiment of the present application can achieve the dual effects of "enhancing the fault-sensitive frequency band" and "suppressing the noise frequency band," resolving the problem of conventional methods' inability to adaptively extract key frequency bands and providing reliable feature information for early diagnosis of mechanical faults.

[0058] S140: performing feature enhancement processing on the sample normalized temperature sub-vector corresponding to each sample feature time series data to obtain a temperature vibration enhancement feature.

[0059] Specifically, in the embodiments of the present application, although the temperature time series data is smooth, fault modes such as abnormal thermal distribution of the reactor are manifested as local gradient changes. Conventional technologies use global averaging or sliding window processing, ignoring the spatial dependence between temperature data points, and are unable to effectively capture subtle temperature gradient changes, resulting in a low recognition rate for heat-related faults such as local blockages, and difficulty in detecting local mutation features in smooth data.

[0060] In order to solve the above technical problems, the embodiment of the present application adopts a local polynomial fitting method based on gradient amplitude weighting to determine the temperature vibration enhancement characteristics.

[0061] In one implementation, the sample temperature time series data includes a plurality of sample temperature values ​​corresponding one-to-one to a plurality of sampling moments; the sample normalized temperature subvector includes a plurality of sample temperature vectors corresponding one-to-one to a plurality of sampling moments and a plurality of sample temperature values; S140 includes: steps (8) to (12), details of which are as follows: Step (8): Determine multiple sample temperature windows based on the sample temperature time series data in each sample feature time series data.

[0062] The temperature window set includes: a plurality of sample temperature values ​​corresponding one-to-one to a plurality of consecutive sampling moments.

[0063] Specifically, in the embodiment of the present application, each temperature window set is obtained based on the window size and the window sliding step size, thereby establishing a local analysis unit; wherein, the formula for determining the number of sample temperature windows is as follows: ; Where, Indicates the number of sample temperature windows; Represents the dimension of sample temperature time series data; represents the window size of the sample temperature window, Based on the fault feature scale adjustment, it can usually be set between 32 and 128 to ensure that the sample temperature window covers the local temperature gradient. In actual operation, The value of is set to 64; Indicates a round-down operation; S represents the window sliding step size, which is used to control the window overlap rate. In actual operation, the value of S can be set to 32.

[0064] The first temperature window in the set The temperature window set is expressed as ; ; ; Where, Represents the sample temperature values ​​included in the temperature window set and their time series positions in the sample temperature time series data. For example, if The value of is 5, which means that the corresponding sample temperature value is at the 5th position in the sample temperature time series data; Indicates the The sample temperature value at the first time sequence position among the multiple sample temperature values ​​included in the temperature window set; Indicates the - the sample temperature value at the first time sequence position among the multiple sample temperature values ​​included in a temperature window set; Indicates the The sample temperature value at the last time sequence position among the multiple sample temperature values ​​included in the temperature window set; Indicates the window sliding step size, which is used to control the window overlap rate. In actual operation, Set the value of to 32.

[0065] Step (9): Based on the multiple sample temperature vectors respectively included in each sample normalized temperature sub-vector, determine the difference between the sample temperature vectors corresponding to any two adjacent sampling moments to obtain multiple sample temperature gradient vectors that are in one-to-one correspondence between the multiple sample temperature vectors.

[0066] Specifically, in the embodiment of the present application, a gradient value is obtained based on the values ​​of adjacent normalized temperature points, thereby identifying potential fault areas; wherein, the formula for determining the sample temperature gradient vector is as follows: ; Where, For the The sample temperature gradient vector corresponding to the sample temperature vector; represents the first in the sample normalized temperature subvector sample temperature vectors; represents the first in the sample normalized temperature subvector -1 sample temperature vector.

[0067] Sample temperature gradient vector The use of backward differencing rather than central differencing is determined by the temperature field characteristics of the green ammonia reactor. The spatial arrangement density of the reactor wall temperature sensors is 5 cm / point. The temperature gradient change caused by the leakage fault only spans 2 to 3 sensors in the initial stage. The central difference will blur the boundaries of these microscale mutations due to the smoothing effect, while the backward difference can amplify the gradient extreme points at the fault edge while maintaining computational efficiency.

[0068] Step (10): Determine the sample gradient weight corresponding to each sample temperature gradient vector based on the sample temperature gradient vector and the preset sensitivity parameter.

[0069] Specifically, in the embodiment of the present application, a gradient weight is obtained based on the gradient value and the sensitivity parameter, thereby strengthening the high gradient area; wherein, the formula for determining the sample gradient weight is as follows: ; Where, Indicates the The gradient weight corresponding to the sample temperature vector; Represents the sensitivity parameter, which is used to control the weight decay rate. In actual operation, The value of is set to 0.5; represents the natural exponential function.

[0070] It should be noted that conventional methods often directly take the absolute value or square of the gradient as the weight, but the normal thermal distribution of the green ammonia reactor has a background gradient of ±0.3℃ / cm, and the fault gradient usually ranges from 1 to 3℃ / cm. Therefore, in the embodiment of the present application, the gradient weight is During the calculation process, the gradient amplitude and weight attenuation are asymmetric, forming a weight separation effect of the fault area / background area. The nonlinear attenuation automatically weakens the fitting weight of the flat area in the polynomial fitting, forcing the regression curve to stick to the high gradient area.

[0071] Step (11): Based on the multiple sample temperature vectors, multiple sample gradient weights and the minimization objective function corresponding to each sample temperature window, determine the sample polynomial coefficients corresponding to each sample temperature window under the constraint of the minimization objective function.

[0072] Specifically, in the embodiment of the present application, polynomial coefficients are obtained based on the normalized temperature points and gradient weights within the window, thereby highlighting the fault indication area; wherein the objective function for highlighting the fault indication area is as follows: - ; Where, express The polynomial order The polynomial coefficients of order are obtained by solving the weighted least squares problem for the objective function, which characterize the temperature polynomial trend of each sample temperature window; ; express of power; Represents the polynomial coefficients The minimization objective function.

[0073] It should be noted that conventional polynomial fitting treats all points in the window equally, but fault-related temperature anomalies are often concentrated in a local area. Therefore, the embodiment of the present application uses gradient weights to Fault-oriented adaptive fitting is achieved. In the low-temperature area caused by blockage, the high weight drives the regression curve concave, and in the high-temperature area of ​​leakage, it is convex. When the polynomial order is When set to 2, the second-order polynomial coefficients The change in the sign of can directly indicate the type of fault, with a negative value indicating blockage and a positive value indicating leakage, which is impossible to achieve using conventional methods.

[0074] Step (12): Determine the sample temperature enhancement feature based on the sample polynomial coefficients corresponding to each sample temperature window.

[0075] Specifically, in the embodiment of the present application, the sample temperature enhancement feature after dimensionality reduction is obtained based on the polynomial coefficients of all orders of all sample temperature windows and the second dimensionality reduction projection matrix; wherein, the sample temperature enhancement feature is defined as ; Where, Indicates the The first sample temperature window The polynomial coefficients of order, Indicates the The first sample temperature window The polynomial coefficients of order, Indicates the The first sample temperature window The polynomial coefficients of order, Indicates the The first sample temperature window The polynomial coefficients of order, Indicates the The first sample temperature window Order polynomial coefficients; Represents vector transpose; Represents the second dimensionality reduction projection matrix, the second dimensionality reduction projection matrix is the model parameter trained in the training process of the fault mode recognition model; the second dimension reduction projection matrix The polynomial coefficients of all orders of all sample temperature windows are initialized using principal component analysis. Orthogonality must be met to retain the maximum variance, dimensionality reduction is achieved to reduce redundancy, and the main trend features are extracted.

[0076] It should be noted that the weighted least squares and the second dimension reduction projection matrix The synergistic effect enables principal component analysis to automatically focus on the feature variance of the weight-enhanced area.

[0077] For example, we use thermal maps to analyze the effect of temperature gradient weighted feature extraction, and compare the detection capabilities of different temperature feature extraction methods for local blockage faults through thermal maps. Figure 6 As shown, Figure 6 This is a comparison chart of the detection effect of the temperature-enhanced feature extraction method provided in the embodiment of the present application on local blockage. Figure 6 The figure includes three sub-graphs, with the horizontal axis representing the reactor length (cm) and the vertical axis representing the reactor circumference (cm). The color mapping of the sub-graphs represents the intensity of the temperature feature (dimensionless relative value). (a) shows the performance of the sliding average method in detecting low-temperature zones caused by local blockages, (b) shows the performance of the global polynomial fitting method in detecting low-temperature zones caused by local blockages, and (c) shows the performance of the gradient weighted method of this technology in detecting low-temperature zones caused by local blockages.

[0078] Figure 6 The performance of three methods, namely sliding average, global polynomial fitting and gradient weighting of this technology (temperature feature enhancement method provided in the embodiment of this application), in detecting low-temperature areas caused by local blockages was compared; the experimental results showed that the thermal map of the sliding average method showed over-smoothing characteristics, and the low-temperature features with a length of 15-25 cm and a circumference of 25-35 cm were severely weakened; although the thermal map of the global polynomial fitting method reflected the overall trend, it completely failed to capture the local blockage characteristics; the thermal map of the temperature feature enhancement method provided in the embodiment of this application clearly showed the feature enhancement effect of the low-temperature fault area, verifying the core value of the gradient weight calculation. The asymmetric weight distribution constructed through the exponential decay function enables the high-gradient area (fault feature area) to obtain a higher fitting weight, thereby accurately capturing the local temperature mutation that the traditional method cannot detect in the polynomial regression.

[0079] S150: Determine a sample dimension reduction temperature-vibration fusion feature vector according to the sample vibration enhancement feature and the sample temperature enhancement feature corresponding to each sample feature time series data.

[0080] Specifically, the sample vibration enhancement features and the sample temperature enhancement features have different dimensions and statistical characteristics. Directly connecting the two features will lead to input redundancy and dimensionality mismatch. Conventional techniques such as principal component analysis or simple feature splicing ignore the interaction between the sample vibration enhancement features and the sample temperature enhancement features. High-dimensional input increases the training complexity of deep neural networks and cannot effectively utilize the complementarity between the two features.

[0081] To solve the above technical problems, the embodiment of the present application adopts a dual-branch attention fusion mechanism to achieve feature fusion and dimensionality reduction input.

[0082] In one implementation, S150 includes steps (13) to (16), the details of which are as follows: Step (13): Determine the sample vibration feature attention weight vector based on the current vibration feature attention parameter matrix and the sample vibration enhancement features corresponding to each sample feature time series data.

[0083] Specifically, in the embodiment of the present application, a sample vibration feature attention weight vector is obtained based on the current vibration feature attention parameter matrix and the sample vibration enhancement feature to strengthen the fault sensitive component; wherein, the formula for determining the sample vibration feature attention weight vector is as follows: ; in, Represents the attention weight vector of the sample vibration feature; Represents the current vibration feature attention parameter matrix, the current vibration feature attention parameter matrix are model parameters trained during the training process of the fault mode recognition model; Indicates the enhanced vibration characteristics of the sample; express Activation function.

[0084] Step (14): Determine the sample temperature feature attention weight vector based on the current temperature feature attention parameter matrix and the sample temperature enhancement features corresponding to each sample feature time series data.

[0085] Specifically, in the embodiment of the present application, a sample temperature feature attention weight vector is obtained based on the current temperature feature attention parameter matrix and the sample temperature enhancement feature to enhance the fault sensitive component; wherein, the formula for determining the sample temperature feature attention weight vector is as follows: ; in, Represents the sample temperature feature attention weight vector; Represents the current temperature feature attention parameter matrix, the current temperature feature attention parameter matrix are model parameters trained during the training process of the fault mode recognition model; represents the temperature enhancement characteristics of the sample; express Activation function.

[0086] Step (15): Determine the sample temperature-vibration fusion feature vector according to the sample vibration enhancement feature, the sample vibration feature attention weight vector, the sample temperature enhancement feature and the sample temperature feature attention weight vector corresponding to each sample feature time series data.

[0087] Specifically, in the embodiment of the present application, a temperature-vibration fusion feature vector is obtained based on element-wise multiplication and feature vectors, thereby retaining modal interaction information; wherein, the formula for determining the sample temperature-vibration fusion feature vector is as follows: ; Where, represents the sample temperature-vibration fusion feature vector; Represents a vector concatenation operation; Represents element-wise multiplication, which is used to scale feature elements by their weights.

[0088] It should be noted that the overall temperature rise or temperature drop trend component in the sample temperature enhancement feature is statistically correlated with the broadband energy in the sample vibration enhancement feature, but the fault sensitive information of the local gradient or specific frequency band energy is unrelated to each other. and the sample vibration feature attention weight vector The cross-modal attention screening enables important feature elements to be scaled to the interval [0, 2], realizing the construction of fault subspace projection in the feature space. When the bearing is worn, the high-frequency vibration energy weight and the local temperature gradient weight are synchronously enhanced, so that the fusion vector forms a joint activation in the corresponding dimension, which is impossible to achieve with the conventional splicing method that does not distinguish the importance of features.

[0089] Step (16): Determine the sample dimensionality-reduced temperature-vibration fusion feature vector according to the sample temperature-vibration fusion feature vector corresponding to each sample feature time series data and the preset third dimensionality-reduction projection matrix.

[0090] Specifically, in the embodiment of the present application, the sample reduced-dimensional temperature-vibration fusion feature vector is obtained based on the sample temperature-vibration fusion feature vector and the third reduced-dimensionality projection matrix, thereby generating a compact input representation, which is expressed as: ; Where, Represents the sample dimension reduction temperature-vibration fusion feature vector; represents the sample temperature-vibration fusion feature vector; Represents the third dimensionality reduction projection matrix, the third dimensionality reduction projection matrix In order to train the model parameters in the fault mode recognition model training process, the sample temperature-vibration fusion feature vector is used and initialized using principal component analysis to reduce the training complexity and enhance feature complementarity.

[0091] S160: Iteratively train the fault pattern recognition model through the dimensionality reduction temperature-vibration fusion feature vectors of multiple samples.

[0092] Specifically, the failure modes of green ammonia reactors are complex, with multiple types of faults such as leakage and blockage coexisting. Conventional deep neural networks use fixed activation functions such as ReLU, which are prone to gradient vanishing or saturation problems in deep networks. They are unable to adaptively learn the abstract representation of complex fault modes, resulting in insufficient recognition ability for complex fault modes.

[0093] Without solving the above technical problems, the embodiments of the present application adopt parameterized exponential linear units in the forward propagation of deep neural networks and combine them with inter-layer feature scaling to achieve adaptive deep feature learning of deep neural networks.

[0094] In one implementation, the fault pattern recognition model includes: a feature extraction module, the feature extraction module includes: a plurality of network layers connected in sequence; S160 includes: steps (17) to (20), the details of which are as follows: Step (17): for each sample dimensionality reduction temperature-vibration fusion feature vector, the sample dimensionality reduction temperature-vibration fusion feature vector is processed in turn through multiple network layers remaining after removing the last network layer in the feature extraction module to obtain the first sample output feature vector.

[0095] Step (18): Processing the sample output feature vector according to the current weight matrix and current bias corresponding to the last network layer in the feature extraction module through the feature extraction module to obtain a second sample output feature vector.

[0096] Specifically, in the embodiment of the present application, the second sample output feature vector is obtained based on the current weight matrix, the current bias, and the first sample output feature vector; wherein the formula for determining the second sample output feature vector is as follows: ; Where, Indicates the last ( A second sample output feature vector output by the network layer; Indicates the first The first sample output feature vector output by the network layer, for the first layer, Represents the sample dimension reduction temperature vibration fusion feature vector ; Indicates the first The current weight matrix corresponding to the network layer; Indicates the first The current bias, current weight matrix and current bias corresponding to each network layer are all model parameters trained during the training process of the fault mode recognition model.

[0097] Step (19): Determine, through the feature extraction module, a sample scaling factor of the last network layer in the feature extraction module based on the first sample output feature vector and the dimension of the first sample output feature vector.

[0098] Specifically, in the embodiment of the present application, based on the second sample output feature vector Get the first The output of the activation function of the network layer, thereby enhancing the nonlinear expression ability; among them, determining the The output formula of the activation function of each network layer is as follows: ; Where, Indicates the The output of the activation function of each network layer; represents the parameterized exponential linear unit activation function; Represents element-wise multiplication; The parameter vector represents the parameterized exponential linear unit. is the model parameter trained in the training process of the fault mode recognition model, the parameter vector Used to control the curvature of negative areas.

[0099] In an embodiment of the present application, a sample scaling factor is obtained based on the norm and dimension of the first sample output feature vector, thereby controlling feature scaling; wherein the formula for determining the sample scaling factor is as follows: ; Where, Indicates the first Sample scaling factor for each network layer; Represents the first sample output feature vector Dimensions; Represents a logarithmic function, with the default base being a natural constant.

[0100] Step (20): Determine the third sample output feature vector based on the second sample output feature vector and the sample scaling factor through the feature extraction module.

[0101] Specifically, in the embodiment of the present application, based on the The output of the activation function of the network layer Hedi Sample scaling factor for each network layer The third sample output feature vector is obtained, thereby completing feature scaling; wherein, the formula for determining the third sample output feature vector is as follows: ; Where, Indicates the first The third sample output feature vector output by the network layer.

[0102] It should be noted that the normalization of the conventional network layer forces zero mean and unit variance, but the fault features may have pattern-specific scale distribution in the feature extraction module, such as the leakage feature norm is greater than the blockage, and the sample scaling factor To solve the problem of feature scale drift in feature extraction module, a mild automatic scaling is constructed by logarithmic function. hour, Provides sub-linear growth, such as hour, , which can be used in scenarios where residual networks are prone to feature explosion. The growth rate approaches , forming a direct stabilization valve.

[0103] In one implementation, the fault pattern recognition model further includes: a feature classification module; after step (20), the method further includes: steps (21) to (24), the details of which are as follows: Step (21): Determine the first sample prediction probability corresponding to each fault mode category according to the current deep feature classification weight vector corresponding to each fault mode category and the third sample output feature vector through the feature classification module.

[0104] Specifically, in the embodiment of the present application, the first The third sample output feature vector output by the network layer contains high-level abstract features, but may weaken the local sensitive patterns in the original sensor data, such as small temperature mutation points. The conventional Softmax classification layer directly uses the third sample output feature vector for linear mapping, ignoring the sample reorganization vector. This lacks detailed information in the image, resulting in insufficient response to early weak faults such as micro-leakage.

[0105] To solve the above technical problems, the embodiment of the present application adopts residual feature bridging and gated confidence fusion mechanism to achieve feature classification by fusing deep abstract features with shallow detail information.

[0106] In an embodiment of the present application, the first sample prediction probability is obtained based on the third sample output feature vector and the current depth feature classification weight vector; wherein the formula for determining the first sample prediction probability is as follows: ; Where, Indicates that it corresponds to Failure mode category The first sample prediction probability of each failure mode category; Indicates the first The third sample output feature vector output by the network layer; Indicates the corresponding The current deep feature classification weight vector of the fault mode categories; Indicates the corresponding The current deep feature classification weight vector of the fault mode category; Failure mode categories and The fault mode categories are different fault mode categories; the current deep feature classification weight vector is the model parameter trained in the training process of the fault mode recognition model.

[0107] It should be noted that the first sample prediction probability In synergy with the residual path, the third sample in the deep layer outputs the feature vector Although it contains high-level abstract patterns, it has a weak response to local details of early faults. The full dimension of the image is obtained instead of the conventional dimensionality reduction before Softmax, so that the subsequent residual gating can accurately locate the details lost by the deep network. When the deep feature response of the micro-leakage fault is weak, the residual path can extract the burr features in the normalized sample. The two complement each other and significantly improve the early fault recall rate.

[0108] Step (22): Determine the sample residual feature vector based on the sample normalized temperature subvector, the sample normalized vibration subvector and the current residual projection matrix through the feature classification module.

[0109] Specifically, in the embodiment of the present application, the sample reorganization vector is extracted by the residual path. The fault-sensitive details in the sample are used to determine the sample residual feature vector, thereby extracting the fault-sensitive details. The formula for determining the sample residual feature vector is as follows: ; Where, represents the sample residual eigenvector; Represents the ReLU activation function; Represents the current residual projection matrix, which is the model parameter trained during the training process of the fault mode recognition model; Represents the sample reorganization vector with a dimension of 4096.

[0110] Step (23): Determine the sample gating weight vector based on the sample residual feature vector and the current gating parameter matrix through the feature classification module.

[0111] Specifically, in the embodiment of the present application, a sample gating weight vector is obtained based on the sample residual feature vector and the current gating parameter matrix, thereby screening key features; wherein, the formula for determining the sample gating weight vector is as follows: ; Where, Indicates the corresponding The sample gating weight vector of the failure mode categories; Indicates the corresponding The current gating parameter matrix of the fault mode category is the model parameters trained in the training process of the fault mode recognition model.

[0112] Step (24): Determine the second sample prediction probability corresponding to each fault mode category based on the current gating parameter matrix, the sample residual feature vector corresponding to each fault mode category, and the sample gating weight vector through the feature classification module.

[0113] Specifically, in the embodiment of the present application, based on the corresponding The sample gating weight vector of the fault mode category and the current residual classification weight vector are obtained corresponding to the first The second sample prediction probability of the fault mode category is determined to fuse the gating mechanism; wherein, the second sample prediction probability corresponding to the first The formula for the second sample predicted probability of a failure mode category is as follows: ; Where, Represents the residual feature corresponding to the The second sample prediction probability of each failure mode category; Indicates the corresponding The sample gating weight vector of the failure mode categories; Indicates the The current residual classification weight vector for the class; Indicates the The current residual classification weight vector of the class is the model parameter trained during the training process of the fault mode recognition model.

[0114] It should be noted that conventional residual connections directly fuse features, while the embodiment of the present application learns an independent sample gating weight vector for each type of fault mode category. For example, for leakage faults, the sample gating weight vector automatically strengthens the dimension corresponding to the gradient sudden change point in the temperature sub-vector. For bearing wear, the gating weight vector focuses on the high-frequency band of the vibration sub-vector. Based on this, when the input sample contains multiple types of fault features at the same time, the sample gating weight vector enables each type of probability calculation to focus only on its sensitive dimensions, avoiding misjudgment caused by feature mutual exclusion.

[0115] In one implementation, the fault pattern recognition model further includes: a feature classification module; after step (24), the method further includes: step (25), the details of which are as follows: Step (25): determining, by a feature classification module, a third sample prediction probability corresponding to each failure mode category based on the first sample prediction probability and the second sample prediction probability corresponding to each failure mode category; The third sample predicted probability indicates a failure mode category of the green ammonia reactor.

[0116] Specifically, in the embodiment of the present application, the final prediction probability, the third sample prediction probability, is obtained based on the first sample prediction probability and the second sample prediction probability, thereby integrating deep abstraction and shallow detail information; wherein, the formula for determining the third sample prediction probability is as follows: ; ; ; Where, Indicates the predicted probability based on the first sample The corresponding confidence level for each failure mode category; Indicates the predicted probability based on the first sample The corresponding confidence level for each failure mode category; Indicates the predicted probability based on the second sample The corresponding confidence level for each failure mode category; Indicates the predicted probability based on the second sample The corresponding confidence level for each failure mode category; Indicates the corresponding The third sample predicted probability of the failure mode category.

[0117] In one implementation, after step (25), the method further includes steps (26) to (29), the details of which are as follows: Step (26): Determine the sample frequency weight corresponding to each failure mode category based on the number of sample feature time series data corresponding to each failure mode category.

[0118] Specifically, the sample distribution of fault mode categories such as normal, leakage, and blockage is often unbalanced. The conventional cross-entropy loss function ignores the weight problem caused by the difference in the number of samples of different categories, resulting in low recognition rate for minority classes such as severe faults, and cannot effectively deal with the dual challenges of category frequency differences and classification difficulty differences.

[0119] To solve the above technical problems, the embodiment of the present application adopts a weighted focal loss function to integrate the sample frequency weight, sample focal modulation term, label, and the predicted probability of the third sample, thereby enhancing the model's sensitivity to rare faults. Before calculating the loss function value using the weighted focal loss function, it is necessary to first calculate the sample frequency weight corresponding to each fault mode category. The formula for determining the sample frequency weight is as follows: ; Where, Indicates the number of sample feature time series data; express The label of the sample feature time series data is The number of sample feature time series data for each failure mode category; ; represents the number of failure mode categories; Indicates the corresponding The sample frequency weights for each failure mode category.

[0120] Step (27): Determine the sample difficulty factor corresponding to each fault mode category based on the sample frequency weight and historical classification accuracy corresponding to each fault mode category.

[0121] Specifically, in the embodiment of the present application, the sample difficulty factor is determined based on the sample frequency weight and the historical classification accuracy, thereby reflecting the classification difficulty; wherein, the formula for determining the sample difficulty factor is as follows: - ; Where, Indicates the corresponding Sample difficulty factor for each failure mode category; Indicates the corresponding The historical classification accuracy of the failure mode categories, Determined based on the last training round; Indicates adjustment parameters, which are used to control the difficulty sensitivity. In actual operation, The value of is set to 2.

[0122] Step (28): Determine the sample focus modulation item corresponding to each fault mode category according to the third sample prediction probability and the sample difficulty factor corresponding to each fault mode category.

[0123] Specifically, in the embodiment of the present application, a sample focus modulation term is obtained based on the third sample prediction probability and the sample difficulty factor, thereby adaptively adjusting the loss weight; wherein, the formula for determining the sample focus modulation term is as follows: ; Where, Indicates the corresponding Sample focus modulation terms for each failure mode category; Indicates the corresponding The third sample predicted probability of the failure mode category.

[0124] Step (29): Determine the loss function value according to the third sample prediction probability, sample frequency weight and sample focus modulation item corresponding to each fault mode category.

[0125] Among them, the loss function value is used to update the model parameters of the fault mode recognition model.

[0126] Specifically, in the embodiment of the present application, the sample frequency weight, the sample focus modulation term, the label, and the third sample prediction probability are used to obtain a loss function value, thereby enhancing sensitivity to rare faults; wherein, the formula for calculating the loss function value is as follows: ; Where, Represents the loss function value calculated based on the weighted focal loss function; Indicates the corresponding One-hot encoding of the labels of the failure mode categories.

[0127] It should be noted that the sample frequency weight Directly combat sample imbalance, increase the loss weight of rare faults, and increase the sample difficulty factor Dynamically perceive the difficulty of classification, for categories with low historical accuracy, such as confusion between blockage and scaling, Increase the sample focus modulation term More aggressively suppress the loss contribution of easy-to-separate samples and the sample focus modulation term The exponential form focuses on difficult samples. hour , the loss of easily separable samples is attenuated, and hour , the loss of difficult samples is retained, the coupling of triple regulation produces fault difficulty adaptive learning, and the loss is and Double suppression, for low sample size hard to distinguish faults, Zoom in and maintain.

[0128] In an embodiment of the present application, during the iterative training of the fault pattern recognition model, a small batch gradient descent strategy is used for iterative optimization; wherein, in each round of iteration, 256 sample time series feature data are randomly sampled from the training set to form a batch of data, and the forward propagation is performed in sequence to calculate the prediction probability, the loss function value is calculated based on the weighted focal loss function, and the gradient of the loss function for all trainable parameters of the model is calculated through the back propagation algorithm; the parameters are updated using an adaptive moment estimation optimizer, and each time a training cycle is completed, the weighted F1 score is calculated on the validation set as a monitoring indicator of the early stopping strategy.

[0129] In an embodiment of the present application, the training termination conditions for iterative training of the fault mode recognition model include two mechanisms. The first is early stopping based on performance convergence. Specifically, when the weighted F1 score of the verification set has not improved for 10 consecutive training cycles (improvement threshold <0.001), the current optimal model parameters are saved and training is terminated; the first is forced stopping at the maximum number of iterations. Specifically, the preset upper limit of the training cycle is 200 rounds. If early stopping is still not triggered when the cycle upper limit is reached, the parameters corresponding to the cycle with the highest F1 score of the verification set are selected as the final model.

[0130] In practice, the loss curve and validation metrics are recorded for each cycle during training. When the loss function value reaches NaN (Not a Number, a special symbol indicating an invalid value) or the validation set performance drops by more than 20%, it indicates a model crash. Training is immediately interrupted and the parameters are rolled back to the previous cycle to reinitialize the optimizer state.

[0131] Distance description, now we conduct a comparative analysis of the fault identification accuracy to evaluate the performance of the overall fault identification system, such as Figure 7 As shown, Figure 7 A comparison chart of different fault mode identification methods provided in the embodiments of this application, Figure 7 The horizontal axis is the label corresponding to the 7 fault identification categories. Figure 7 The vertical axis is the classification accuracy (dimensionless), Figure 7The recognition performance of SVM (Support Vector Machine), random forest, conventional CNN (Convolutional Neural Network) and the present technology (the green ammonia reactor fault pattern recognition method provided in the embodiment of the present application) was compared.

[0132] according to Figure 7 It can be seen that conventional methods such as SVM, random forest, and conventional CNN perform the weakest on "rotor imbalance" and "bearing wear", with classification accuracy rates lower than 0.85, and fail to handle multimodal feature heterogeneity. The embodiments of the present application have the highest classification accuracy rates in all fault mode categories, especially for "rotor imbalance" and "bearing wear". The experimental results show that: 1) group normalization retains cross-modal feature complementarity; 2) the time-frequency / temperature feature enhancement mechanism extracts weak fault signals; 3) dual-branch attention fusion enables the coordinated activation of temperature gradient weights and vibration high-frequency weights.

[0133] S220: performing feature enhancement processing and feature fusion processing on the target vibration time series data and the target temperature time series data in sequence to obtain a target dimension-reduced temperature-vibration fusion feature vector.

[0134] Specifically, in the embodiment of the present application, the specific implementation details of S220 can be implemented with reference to steps (13) to (16), etc., and will not be repeated here.

[0135] S230: According to the target dimension-reduced temperature-vibration fusion feature vector, a pre-trained fault mode recognition model is used to determine the fault mode category corresponding to the target feature time series data.

[0136] Specifically, in an embodiment of the present application, when the target dimensionality reduction temperature-vibration fusion feature vector is input into a pre-trained fault mode recognition model, the pre-trained fault mode recognition model can output the predicted fault mode category corresponding to the target feature time series data.

[0137] Second, the present application provides a green ammonia reactor fault mode identification device, such as Figure 8 As shown, Figure 8 This is a structural diagram of a green ammonia reactor fault pattern recognition device provided in an embodiment of the present application, the device comprising: a data acquisition module 310, a data processing module 320 and a prediction module 330; The data acquisition module 310 is used to obtain target characteristic time series data of the target green ammonia reactor within a target time period; wherein the target characteristic time series data includes: target vibration time series data and target temperature time series data; The data processing module 320 is used to perform feature enhancement processing and feature fusion processing on the target vibration time series data and the target temperature time series data in sequence to obtain a target dimension-reduced temperature-vibration fusion feature vector; The prediction module 330 is used to determine the fault mode category corresponding to the target feature time series data through a pre-trained fault mode recognition model according to the target dimensionality-reduced temperature-vibration fusion feature vector.

[0138] In one implementation, the apparatus includes: a training module; A training module is used to obtain multiple sample feature time series data of a sample green ammonia reactor within a sample time period; wherein the sample feature time series data includes: sample vibration time series data and sample temperature time series data; The training module is further configured to perform normalization processing on the sample temperature time series data and the sample vibration time series data in each sample feature time series data according to the noise characteristics corresponding to the vibration time series data and the temperature time series data of the green ammonia reactor, respectively, to obtain a sample normalized temperature subvector and a sample normalized vibration subvector; The training module is further used to perform feature enhancement processing on the sample normalized vibration sub-vector corresponding to each sample feature time series data to obtain the sample vibration enhancement feature; The training module is further used to perform feature enhancement processing on the sample normalized temperature subvector corresponding to each sample feature time series data to obtain a sample temperature enhancement feature; The training module is further used to determine the sample dimensionality reduction temperature-vibration fusion feature vector according to the sample vibration enhancement feature and the sample temperature enhancement feature corresponding to each sample feature time series data; The training module is also used to iteratively train the fault pattern recognition model through multiple sample dimension reduction temperature-vibration fusion feature vectors.

[0139] In one implementation, the sample vibration time series data includes a plurality of sample acceleration composite amplitudes corresponding one-to-one to a plurality of sampling moments; the training module is further configured to perform clustering processing on the sample vibration time series data based on the amplitudes of the sample vibration time series data at each of the sample acceleration composite amplitudes having different values ​​in each of the sample feature time series data to obtain the number of sample frequency bands corresponding to the sub-frequency bands of each sample normalized vibration sub-vector; The training module is further used to perform wavelet packet decomposition processing on each sample normalized vibration sub-vector according to the number of sample frequency bands to obtain a sample wavelet packet frequency band coefficient vector of each sub-band; The sample wavelet packet frequency band coefficient vector includes: a sample wavelet packet coefficient vector that corresponds one-to-one to a plurality of vibration frequencies included in the sub-frequency band; The training module is further used to determine the sample adaptive weight corresponding to each sub-band according to the multiple sample wavelet packet coefficient vectors corresponding to each sub-band; The training module is further used to determine the sample vibration enhancement feature according to the sample wavelet packet frequency band coefficient vector and the sample adaptive weight corresponding to each sub-band.

[0140] In one implementation, the sample temperature time series data includes a plurality of sample temperature values ​​corresponding one-to-one to a plurality of sampling moments; the sample normalized temperature subvector includes a plurality of sample temperature vectors corresponding one-to-one to the plurality of sampling moments and the plurality of sample temperature values; the training module is further configured to determine a plurality of sample temperature windows based on the sample temperature time series data in each sample feature time series data; wherein the temperature window set includes: a plurality of sample temperature values ​​corresponding one-to-one to a plurality of consecutive sampling moments; The training module is further configured to determine, based on the multiple sample temperature vectors respectively included in each sample normalized temperature sub-vector, a difference between the sample temperature vectors corresponding to any two adjacent sampling moments, and obtain multiple sample temperature gradient vectors in one-to-one correspondence with the multiple sample temperature vectors; The training module is further configured to determine a sample gradient weight corresponding to each sample temperature gradient vector based on the sample temperature gradient vector and a preset sensitivity parameter; The training module is further configured to determine, based on a plurality of sample temperature vectors, a plurality of sample gradient weights, and a minimization objective function corresponding to each sample temperature window, a sample polynomial coefficient corresponding to each sample temperature window under the constraint of the minimization objective function; The training module is further used to determine the sample temperature enhancement feature according to the sample polynomial coefficients corresponding to each sample temperature window.

[0141] In one implementation, the training module is further configured to determine a sample vibration feature attention weight vector based on the current vibration feature attention parameter matrix and the sample vibration enhancement features corresponding to each sample feature time series data; The training module is further configured to determine a sample temperature feature attention weight vector based on the current temperature feature attention parameter matrix and the sample temperature enhancement features corresponding to each sample feature time series data; The training module is further used to determine the sample temperature-vibration fusion feature vector according to the sample vibration enhancement feature, the sample vibration feature attention weight vector, the sample temperature enhancement feature and the sample temperature feature attention weight vector corresponding to each sample feature time series data; The training module is also used to determine the sample dimensionality reduction temperature-vibration fusion feature vector based on the sample temperature-vibration fusion feature vector corresponding to each sample feature time series data and the preset third dimensionality reduction projection matrix.

[0142] In one implementation, the fault pattern recognition model includes: a feature extraction module, the feature extraction module including: a plurality of sequentially connected network layers; a training module further configured to process, for each sample dimension-reduced temperature-vibration fusion feature vector, the sample dimension-reduced temperature-vibration fusion feature vector in sequence through the plurality of network layers remaining after excluding the last network layer in the feature extraction module to obtain a first sample output feature vector; The training module is further configured to process the sample output feature vector according to the current weight matrix and the current bias corresponding to the last network layer in the feature extraction module through the feature extraction module to obtain a second sample output feature vector; The training module is further configured to determine, through the feature extraction module, a sample scaling factor of a last network layer in the feature extraction module based on the first sample output feature vector and the dimension of the first sample output feature vector; The training module is further configured to determine, through the feature extraction module, a third sample output feature vector according to the second sample output feature vector and a sample scaling factor.

[0143] In one implementation, the fault mode recognition model further includes: a feature classification module; a training module further configured to determine, through the feature classification module, a first sample prediction probability corresponding to each fault mode category based on the current deep feature classification weight vector corresponding to each fault mode category and the third sample output feature vector; The training module is further configured to determine a sample residual feature vector according to the sample normalized temperature subvector, the sample normalized vibration subvector, and the current residual projection matrix through the feature classification module; The training module is further used to determine the sample gating weight vector according to the sample residual feature vector and the current gating parameter matrix through the feature classification module; The training module is also used to determine the second sample prediction probability corresponding to each fault mode category based on the current gating parameter matrix, the sample residual feature vector corresponding to each fault mode category and the sample gating weight vector through the feature classification module.

[0144] In one implementation, the training module is further configured to determine, through the feature classification module, a third sample prediction probability corresponding to each failure mode category based on the first sample prediction probability and the second sample prediction probability corresponding to each failure mode category; The third sample predicted probability indicates a failure mode category of the green ammonia reactor.

[0145] In one implementation, the training module is further configured to determine a sample frequency weight corresponding to each failure mode category based on the number of samples corresponding to each failure mode category in the plurality of sample feature time series data; The training module is further used to determine the sample difficulty factor corresponding to each failure mode category based on the sample frequency weight and historical classification accuracy corresponding to each failure mode category; The training module is further configured to determine a sample focus modulation item corresponding to each failure mode category based on the third sample prediction probability and the sample difficulty factor corresponding to each failure mode category; The training module is further configured to determine a loss function value based on the third sample prediction probability, sample frequency weight, and sample focus modulation term corresponding to each fault mode category; Among them, the loss function value is used to update the model parameters of the fault mode recognition model.

[0146] Third, the present application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, steps S210 to S230 provided in the above embodiment are implemented.

[0147] Fourth, the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, steps S210 to S230 of the above embodiment are executed.

[0148] Fifth, the computer program product provided in this application includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the method in the previous method embodiment. For specific implementation, please refer to steps S210~S230 of the method embodiment, which will not be repeated here.

[0149] In the embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interface, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0150] In addition, the units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0151] Furthermore, the functional modules in each embodiment of the present application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0152] It should be noted that if the function is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the existing technology, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program code.

[0153] In this document, relational terms such as first and second, etc. are used merely to distinguish one entity or operation from another entity or operation, but do not necessarily require or imply any actual relationship or order between these entities or operations.

[0154] The above description is merely an embodiment of the present application and is not intended to limit the scope of protection of the present application. For those skilled in the art, various modifications and variations of the present application are possible. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A method for identifying fault patterns of a green ammonia reactor, characterized in that: The method comprises: Obtain target characteristic time series data of a target green ammonia reactor within a target time period; wherein the target characteristic time series data includes: target vibration time series data and target temperature time series data; For the target vibration time series data and the target temperature time series data, feature enhancement processing and feature fusion processing are sequentially performed to obtain a target dimension-reduced temperature-vibration fusion feature vector; According to the target dimension-reduced temperature-vibration fusion feature vector, a fault mode category corresponding to the target feature time series data is determined through a pre-trained fault mode recognition model.

2. The method according to claim 1, characterized in that Before obtaining the target characteristic time series data of the target green ammonia reactor within the target time period, the method further includes: Acquire multiple sample characteristic time series data of the sample green ammonia reactor within a sample time period; wherein the sample characteristic time series data includes: sample vibration time series data and sample temperature time series data; According to the noise characteristics corresponding to the vibration time series data and the temperature time series data of the green ammonia reactor, the sample temperature time series data and the sample vibration time series data in each of the sample feature time series data are respectively normalized corresponding to the noise characteristics to obtain a sample normalized temperature subvector and a sample normalized vibration subvector; Performing feature enhancement processing on the sample normalized vibration sub-vector corresponding to each of the sample feature time series data to obtain a sample vibration enhancement feature; Performing feature enhancement processing on the sample normalized temperature subvector corresponding to each of the sample feature time series data to obtain a sample temperature enhancement feature; Determining a sample dimension reduction temperature-vibration fusion feature vector according to the sample vibration enhancement feature and the sample temperature enhancement feature corresponding to each of the sample feature time series data; The fault mode recognition model is iteratively trained using a plurality of the sample dimension-reduced temperature-vibration fusion feature vectors.

3. The method according to claim 2, characterized in that The sample vibration time series data includes a plurality of sample acceleration composite amplitudes corresponding one-to-one to a plurality of sampling moments; and the feature enhancement processing is performed on the sample normalized vibration sub-vector corresponding to each of the sample feature time series data to obtain the sample vibration enhancement feature, including: performing clustering processing on the sample vibration time series data according to the amplitudes of the sample vibration time series data at different sample acceleration composite amplitudes in each of the sample feature time series data to obtain the number of sample frequency bands corresponding to the sub-frequency bands of each sample normalized vibration sub-vector; According to the number of the sample frequency bands, performing wavelet packet decomposition processing on each of the sample normalized vibration sub-vectors to obtain a sample wavelet packet frequency band coefficient vector of each of the sub-frequency bands; The sample wavelet packet frequency band coefficient vector includes: a sample wavelet packet coefficient vector that corresponds one-to-one to a plurality of vibration frequencies included in the sub-frequency band; Determining a sample adaptive weight corresponding to each of the sub-frequency bands according to a plurality of sample wavelet packet coefficient vectors corresponding to each of the sub-frequency bands; The sample vibration enhancement feature is determined according to the sample wavelet packet frequency band coefficient vectors and the sample adaptive weights corresponding to the respective sub-frequency bands.

4. The method according to claim 2, characterized in that The sample temperature time series data includes a plurality of sample temperature values ​​corresponding one-to-one to a plurality of sampling moments; the sample normalized temperature subvector includes a plurality of sample temperature vectors corresponding one-to-one to the plurality of sampling moments and the plurality of sample temperature values; and the feature enhancement processing is performed on the sample normalized temperature subvector corresponding to each of the sample feature time series data to obtain a sample temperature enhancement feature, including: Determine a plurality of sample temperature windows according to the sample temperature time series data in each of the sample characteristic time series data; wherein the temperature window set includes: a plurality of sample temperature values ​​corresponding one-to-one to a plurality of consecutive sampling moments; Determining, based on the plurality of sample temperature vectors respectively included in each of the sample normalized temperature sub-vectors, a difference between the sample temperature vectors corresponding to any two adjacent sampling moments, to obtain a plurality of sample temperature gradient vectors in one-to-one correspondence with the plurality of sample temperature vectors; Determining a sample gradient weight corresponding to each of the sample temperature gradient vectors according to the sample temperature gradient vectors and a preset sensitivity parameter; Determining, based on the plurality of sample temperature vectors, the plurality of sample gradient weights, and a minimization objective function corresponding to each of the sample temperature windows, the sample polynomial coefficients corresponding to each of the sample temperature windows under the constraints of the minimization objective function; The sample temperature enhancement feature is determined according to the sample polynomial coefficients corresponding to the respective sample temperature windows.

5. The method according to claim 2, characterized in that The step of determining a sample dimensionality reduction temperature-vibration fusion feature vector according to the sample vibration enhancement feature and the sample temperature enhancement feature corresponding to each of the sample feature time series data includes: Determining a sample vibration feature attention weight vector according to the current vibration feature attention parameter matrix and the sample vibration enhancement features corresponding to each of the sample feature time series data; Determine a sample temperature feature attention weight vector according to the current temperature feature attention parameter matrix and the sample temperature enhancement features corresponding to each of the sample feature time series data; Determine a sample temperature-vibration fusion feature vector according to the sample vibration enhancement feature, the sample vibration feature attention weight vector, the sample temperature enhancement feature, and the sample temperature feature attention weight vector corresponding to each of the sample feature time series data; The sample dimensionality-reduced temperature-vibration fusion feature vector is determined according to the sample temperature-vibration fusion feature vector corresponding to each of the sample feature time series data and a preset third dimensionality-reduced projection matrix.

6. The method according to claim 2, characterized in that The fault pattern recognition model includes: a feature extraction module, which includes: a plurality of sequentially connected network layers; the fault pattern recognition model is iteratively trained using the plurality of sample dimension reduction temperature-vibration fusion feature vectors, including: For each of the sample dimension reduction temperature-vibration fusion feature vectors, the sample dimension reduction temperature-vibration fusion feature vectors are processed in sequence through the plurality of network layers remaining after excluding the last network layer in the feature extraction module to obtain a first sample output feature vector; Processing the sample output feature vector by the feature extraction module according to a current weight matrix and a current bias corresponding to a last network layer in the feature extraction module to obtain a second sample output feature vector; determining, by the feature extraction module, a sample scaling factor for a last network layer in the feature extraction module based on the first sample output feature vector and the dimension of the first sample output feature vector; A third sample output feature vector is determined by the feature extraction module according to the second sample output feature vector and the sample scaling factor.

7. The method according to claim 6, characterized in that The fault mode recognition model further includes: a feature classification module; after determining the third sample output feature vector, the method further includes: Determining, by the feature classification module, a first sample prediction probability corresponding to each of the fault mode categories based on the current deep feature classification weight vector corresponding to each of the fault mode categories and the third sample output feature vector; Determining, by the feature classification module, a sample residual feature vector according to the sample normalized temperature subvector, the sample normalized vibration subvector, and a current residual projection matrix; Determining a sample gating weight vector according to the sample residual feature vector and the current gating parameter matrix by the feature classification module; The feature classification module determines the second sample prediction probability corresponding to each of the fault mode categories according to the current gating parameter matrix, the sample residual feature vector corresponding to each of the fault mode categories, and the sample gating weight vector.

8. The method according to claim 7, characterized in that After determining the second sample prediction probability corresponding to each of the failure mode categories, the method further includes: Determining, by the feature classification module, a third sample prediction probability corresponding to each of the failure mode categories based on the first sample prediction probability and the second sample prediction probability corresponding to each of the failure mode categories; The third sample predicted probability indicates a failure mode category of the green ammonia reactor.

9. The method according to claim 7, characterized in that After determining the third sample prediction probability corresponding to each of the failure mode categories, the method further includes: Determining a sample frequency weight corresponding to each of the failure mode categories according to the number of sample feature time series data corresponding to each of the failure mode categories; Determining a sample difficulty factor corresponding to each of the failure mode categories according to the sample frequency weight and historical classification accuracy corresponding to each of the failure mode categories; Determining a sample focus modulation item corresponding to each of the failure mode categories according to the third sample prediction probability and the sample difficulty factor respectively corresponding to each of the failure mode categories; determining a loss function value according to the third sample prediction probability, the sample frequency weight, and the sample focus modulation term corresponding to each of the fault mode categories; The loss function value is used to update the model parameters of the fault mode recognition model.

10. A green ammonia reactor fault pattern recognition device, characterized in that: The device includes: a data acquisition module, a data processing module and a prediction module; The data acquisition module is used to obtain target characteristic time series data of the target green ammonia reactor within a target time period; wherein the target characteristic time series data includes: target vibration time series data and target temperature time series data; The data processing module is used to perform feature enhancement processing and feature fusion processing on the target vibration time series data and the target temperature time series data in sequence to obtain a target dimensionality-reduced temperature-vibration fusion feature vector; The prediction module is used to determine the fault mode category corresponding to the target feature time series data through a pre-trained fault mode recognition model based on the target dimensionality reduction temperature-vibration fusion feature vector.

Images