A method and apparatus for identifying failure modes in a green ammonia reactor

Abstract

This application relates to the field of data processing and data recognition technology, and in particular to a method and apparatus for fault mode recognition of a green ammonia reactor. The method includes: acquiring target feature time-series data of the target green ammonia reactor within a target time period; wherein, the target feature time-series data includes: target vibration time-series data and target temperature time-series data; performing feature enhancement processing and feature fusion processing sequentially on the target vibration time-series data and the target temperature time-series data to obtain a target dimension-reduced temperature-vibration fusion feature vector; and determining the fault mode category corresponding to the target feature time-series data through a pre-trained fault mode recognition model based on the target dimension-reduced temperature-vibration fusion feature vector. This application can solve the technical problem in the prior art that it is difficult to accurately determine the fault mode category of a green ammonia reactor.

Description

Technical Field

[0001] This application relates to the field of data processing and data recognition technology, and in particular to a method and apparatus for identifying failure modes of a green ammonia reactor. Background Technology

[0002] In the field of renewable energy, the green hydrogen industry chain is a series of industrial links formed around the production of green hydrogen from renewable energy sources (such as wind power, photovoltaic, hydropower, etc.), as well as its development, storage, transportation, application and recycling. Among them, the green ammonia reactor is the core equipment in the green hydrogen industry chain, which is used to synthesize ammonia (NH3) from nitrogen (N2) and hydrogen (H2) through renewable energy-driven chemical reactions, so as to achieve "zero carbon" or "low carbon" ammonia production.

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

[0004] In practical applications, although there are multiple failure modes of green ammonia reactors, all of these failure modes initially exhibit small local gradient changes in the temperature field or subtle changes in narrow-band energy in the vibration signal. The signal amplitude is low and is easily masked by environmental noise and process fluctuations. Therefore, existing technologies cannot accurately determine the failure mode category of green ammonia reactors. Summary of the Invention

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

[0006] In a first aspect, this application provides a method for identifying failure modes in a green ammonia reactor, the method comprising:

[0007] Acquire 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;

[0008] For the target vibration time series data and the target temperature time series data, feature enhancement processing and feature fusion processing are performed sequentially to obtain the target dimension-reduced temperature and vibration fusion feature vector;

[0009] Based on the target reduced-dimensional temperature and vibration fusion feature vector, the fault mode category corresponding to the target feature time series data is determined by a pre-trained fault mode recognition model.

[0010] Secondly, this application provides a green ammonia reactor failure mode identification device, the device comprising: a data acquisition module, a data processing module, and a prediction module;

[0011] The data acquisition module is used to acquire the 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;

[0012] 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 the target dimension-reduced temperature and vibration fusion feature vector.

[0013] The prediction module is used to determine the fault mode category corresponding to the target feature time series data based on the target reduced temperature and vibration fusion feature vector through a pre-trained fault mode recognition model.

[0014] Beneficial effects:

[0015] After collecting two types of time-series data, vibration and temperature, this application utilizes the sensitivity of vibration time-series data to mechanical faults 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 due to a single sensor. Therefore, it can solve the problems of weak early-stage fault signals in green ammonia reactors that are easily interfered with, such as temperature gradient changes that may be masked by process thermal fluctuations and vibration narrowband energy fluctuations that are easily mixed with environmental noise.

[0016] This application extracts key features from vibration time series data and temperature time series data through feature enhancement processing, and then fuses them into a low-dimensional feature vector through a dimensionality reduction algorithm. This process not only preserves fault-related features but also eliminates noise interference, enabling early weak anomalies to form distinguishable patterns in the feature space, laying the foundation for subsequent accurate classification. Therefore, it can solve the problem that the original time series data contains a lot of redundant information, making it difficult to extract effective fault features through direct analysis.

[0017] The fault mode recognition model pre-trained in this application can automatically match feature time series data with fault mode categories. Even when faced with variables such as process fluctuations or equipment aging, it can still maintain a high recognition accuracy. This classification method completely solves the technical problem of low amplitude and easy confusion of early fault signals. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. The following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart of a green ammonia reactor failure mode identification method provided in an embodiment of this application;

[0020] Figure 2 Another flowchart of the green ammonia reactor failure mode identification method provided in the embodiments of this application;

[0021] Figure 3 A diagram illustrating the impact of the grouping normalization strategy provided in this application embodiment on feature quality;

[0022] Figure 4 A comparison diagram of the feature enhancement effect of vibration time series data provided in the embodiments of this application;

[0023] Figure 5(a) is a frequency band analysis diagram of the sample vibration time series data provided in the embodiment of this application;

[0024] Figure 5(b) is a frequency band analysis diagram of traditional wavelet packet decomposition provided in the embodiments of this application;

[0025] Figure 5(c) is a frequency band analysis diagram of the vibration feature enhancement method provided in the embodiment of this application;

[0026] Figure 6 The images show a comparison of the detection effects of the temperature enhancement feature extraction method provided in this application embodiment on local blockage; wherein, (a) is a schematic diagram of the performance of the moving average method provided in this application embodiment when detecting the low temperature region caused by local blockage; (b) is a schematic diagram of the performance of the global polynomial fitting method provided in this application embodiment when detecting the low temperature region caused by local blockage; and (c) is a schematic diagram of the performance of the gradient weighting method provided in this application embodiment when detecting the low temperature region caused by local blockage.

[0027] Figure 7 A comparison diagram of different fault mode identification methods provided in the embodiments of this application;

[0028] Figure 8 This is a structural diagram of the green ammonia reactor fault mode identification device provided in an embodiment of this application. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0030] First, this application provides a method for identifying failure modes in a green ammonia reactor, such as... Figure 1 As shown, Figure 1 A flowchart of a green ammonia reactor failure mode identification method provided in this application embodiment is shown below. The method includes steps S210-S230, as detailed below:

[0031] S210: Obtain the target characteristic time series data of the target green ammonia reactor within the target time period.

[0032] The target characteristic time series data includes: target vibration time series data and target temperature time series data.

[0033] Specifically, in the embodiments of this application, the "target green ammonia reactor" is a green ammonia reactor with a "failure mode identification category" requirement; the "target feature time series data" is the basic data required to identify the failure mode category of the target green ammonia reactor.

[0034] 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 in a linear array on the pipe wall of the target green ammonia reactor, with a spatial density of one monitoring point every 5 centimeters. The vibration sensors are installed at the connection points between the bearing support points and the pipelines in the target green ammonia reactor.

[0035] In actual operation, temperature and vibration signals are acquired in real time through an industrial data acquisition system with a sampling frequency of 1024Hz. The target feature time series data is obtained by continuous acquisition for 4 seconds.

[0036] Each target feature time series data contains 4096 dimensions of raw data. The first 2048 dimensions of the target feature time series data are the target temperature time series data, that is, the temperature values ​​collected at 2048 sampling times, arranged in a time sequence. The last 2048 dimensions of the target feature time series data are the target vibration time series data. In actual operation, the target vibration time series data is the composite amplitude of three-dimensional vibration acceleration. In three-dimensional space, the vibration acceleration of an object can be decomposed into three mutually perpendicular components. The composite amplitude is the magnitude of the vector sum of these three components, reflecting the total intensity of the vibration. Therefore, the target vibration time series data is the composite amplitude of acceleration collected at 2048 sampling times, arranged in a time sequence.

[0037] It is important to emphasize that the temperature value and the composite acceleration amplitude are collected synchronously. That is, the 2048 sampling times corresponding to the target vibration time series data and the target temperature time series data should be the same. In other words, at each sampling time, a temperature value and a composite acceleration amplitude are collected simultaneously, and then combined into the target feature time series data.

[0038] In this embodiment of the application, after the target feature time series data is obtained, the target feature time series data is further preprocessed by feature enhancement processing and other methods, and then the preprocessed data is input into the pre-trained fault mode recognition model. The data output by the fault mode recognition model is the fault mode category corresponding to the target feature time series data. The fault mode category indicates the category to which the fault mode of the target green ammonia reactor belongs.

[0039] In one implementation, before S210, such as Figure 2 As shown, Figure 2 Another flowchart of the green ammonia reactor failure mode identification method provided in this application embodiment is shown below. The method further includes steps S110 to S160, as detailed below:

[0040] S110: Obtain time-series data of multiple sample characteristics from the green ammonia reactor within the sample time period. These time-series data include: sample vibration time-series data and sample temperature time-series data.

[0041] Specifically, in the embodiments of this application, the fault mode recognition model needs to be iteratively trained before applying the fault mode recognition model.

[0042] In practice, temperature and vibration signals are acquired in real time through an industrial data acquisition system with a sampling frequency of 1024Hz. A sample feature time series data is obtained by continuously acquiring data for 4 seconds. After continuous acquisition for 3 months, 8000 sample feature time series data are extracted from the sample feature time series data collected over the 3 months. Among them, the number of positive samples and negative samples are both 4000.

[0043] Each sample feature time series data contains 4096 dimensions of raw data. The first 2048 dimensions of the sample feature time series data are the sample temperature time series data, that is, the temperature values ​​collected at 2048 sampling times, arranged in a time sequence. The last 2048 dimensions of the sample feature time series data are the sample vibration time series data. In actual operation, the sample vibration time series data is the composite amplitude of three-dimensional vibration acceleration. In three-dimensional space, the vibration acceleration of an object can be decomposed into three mutually perpendicular components. The composite amplitude is the magnitude of the vector sum of these three components, reflecting the total intensity of the vibration. Therefore, the sample vibration time series data is the composite amplitude of acceleration collected at 2048 sampling times, arranged in a time sequence.

[0044] In this embodiment of the 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 preset labels; wherein, the labeling of the fault mode category is determined based on the green ammonia reactor operation log and expert diagnosis results, and the label indicates the fault mode category.

[0045] In practice, when the sample characteristic time series data indicates that the green ammonia reactor is in normal operation, it is labeled as "normal"; when the sample characteristic time series data indicates that a gas leak is detected, it is labeled as "sealing leak" or "interface leak" depending on the leak location (sealing ring / flange interface); when the sample characteristic time series data indicates that an abnormality is detected in the catalyst bed, it is labeled as "partial blockage" or "global blockage" depending on the pressure difference change; when the sample characteristic time series data indicates that an abnormality is detected in mechanical components, it is labeled as "bearing wear" or "rotor imbalance" depending on the spectral characteristics, forming a total of 7 types of fault labels.

[0046] S120: Based on the noise characteristics corresponding to the vibration time series data and temperature time series data of the green ammonia reactor, the sample temperature time series data and sample vibration time series data in each sample feature time series data are normalized according to the noise characteristics to obtain the sample normalized temperature sub-vector and sample normalized vibration sub-vector.

[0047] 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, while the noise characteristics of the temperature time-series data are smooth and low-noise. The physical characteristics and statistical distributions of the two are significantly different. Therefore, if the conventional global normalization method is used to normalize the vibration time-series data and 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 the fault modes such as leakage or blockage, thereby reducing the accuracy of fault mode category identification.

[0048] To address the aforementioned technical issues, this application employs a grouping normalization strategy. Based on the noise characteristics corresponding to the vibration time series data and temperature time series data of the green ammonia reactor, the heterogeneous characteristics of the temperature time series data and vibration time series data are independently normalized.

[0049] In one implementation, step S120 includes steps (1) to (3), as detailed below:

[0050] Step (1): Based on the sample temperature time series data and sample vibration time series data in each sample feature time series data, determine the sample vibration time series data sub-vector and sample temperature time series data sub-vector corresponding to each sample feature time series data respectively.

[0051] Specifically, in the embodiments of this application, a sample temperature time series data sub-vector is obtained based on the first 2048 dimensions 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 physical scale mismatch problem.

[0052] In practice, a sample temperature time series data subvector is defined. Dimension of sample temperature time series data subvectors Sample vibration time series data subvector Dimension of sample vibration time series data sub-vectors .

[0053] In the formula, This represents the time-series data of the sample green ammonia reactor, with a dimension of 4096, consisting of the sample temperature time-series data sub-vector. and sample vibration time series data subvectors It is pieced together; Representing time series data of sample features The former Dimensional elements; Representing time series data of sample features The Dimension to the Dimensional elements.

[0054] Step (2): Based on the noise characteristics corresponding to the vibration time series data of the green ammonia reactor, normalize the sample vibration time series data sub-vectors to obtain the sample normalized vibration sub-vectors.

[0055] Specifically, a normalized vibration time-series data sub-vector 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 shown below:

[0056] ; ;

[0057] ;

[0058] In the formula, This represents a sample-normalized vibration time series data subvector; It represents the average of the composite amplitudes of multiple accelerations included in the sample vibration time series data; This indicates the sample vibration time series data included The first of the composite acceleration amplitudes The combined amplitude of the acceleration; ; The standard deviation represents the composite amplitude of multiple accelerations included in the sample vibration time series data; This represents a very small constant, used to prevent division by zero errors. In practice, it can be... The value is set to .

[0059] Step (3): Based on the noise characteristics corresponding to the temperature time series data of the green ammonia reactor, normalize the sample temperature time series data subvector to obtain the sample normalized temperature subvector.

[0060] Specifically, a normalized temperature time-series data subvector is obtained based on the 10th and 90th percentiles of the sample temperature time-series data, thereby eliminating the influence of outliers and preserving smooth features; the formula for normalizing the sample temperature time-series data is shown below:

[0061] ;

[0062] In the formula, Represents a sample normalized temperature time series data subvector;

[0063] Indicates the 10th percentile of the sample temperature time series data;

[0064] This represents the 90th percentile of the sample temperature time series data.

[0065] In this embodiment of the application, when the sample normalized vibration time series data subvector is obtained... and sample normalized temperature time series data subvectors Then, you can and The vectors are concatenated to form a sample recombination vector. The dimension is 4096.

[0066] It should be noted that this application's embodiments abandon the conventional global normalization homogenization process and solve the core contradiction of multimodal data fusion through a strategy of independent normalization of heterogeneous features and splicing and recombination. Temperature time series data adopts quantile-based normalization, which essentially uses robust statistics to isolate outliers. Vibration time series data adopts difference normalization with a smoothing factor. The two types of features are processed independently and then spliced, generating a cross-modal feature complementary effect in fault mode recognition. When the reactor is partially blocked, the vibration energy distribution in the temperature gradient rising region will exhibit a specific frequency band attenuation. The recombined sample reconstructed vector This makes the association pattern linearly separable in the feature space.

[0067] For example, a comparative analysis of the effects of grouping normalization strategies is conducted to verify the effectiveness of these strategies in improving the quality of fault characteristics. The grouping normalization strategy involves normalizing the sample vibration time-series data and the sample temperature time-series data separately. Figure 3 As shown, Figure 3 This is a diagram illustrating the impact of the grouping normalization strategy provided in this application embodiment on feature quality. Figure 3 The horizontal axis represents the labels corresponding to the seven fault identification categories. Figure 3 The vertical axis represents the feature discrimination (F1 score), which is a comprehensive metric used in machine learning and statistics to evaluate the performance of classification models.

[0068] Figure 3 This paper compares the feature discrimination of four methods—global normalization, temperature grouping normalization, vibration grouping normalization, and the present technology (combination)—across seven fault modes. The "present technology (combination)" refers to the grouping normalization strategy provided in this application's embodiments. Experimental results show that global normalization (the first column in each group of columns) performs the worst across all fault mode categories, with significantly low feature discrimination. This is because the high-amplitude noise in the vibration time series data overwhelms the subtle changes in the temperature time series data, thus destroying the fault-sensitive features. The grouping normalization strategy provided in this application's embodiments (the last column in each group of columns) achieves the highest feature discrimination across all fault types, particularly showing significant advantages in "interface leakage" and "bearing wear." This indicates that quantile normalization is used to isolate outliers in the temperature time series data, and difference normalization is used to suppress high-frequency noise in the vibration time series data. The two methods are processed independently and then concatenated, making the cross-modal correlation features of temperature gradient changes and vibration energy distribution linearly separable in the feature space.

[0069] S130: Perform feature enhancement processing on the normalized vibration subvectors corresponding to each sample feature time series data to obtain sample vibration enhancement features.

[0070] Specifically, in the embodiments of this application, the sample feature time series data contains high-frequency components and transient impact signals that are strongly correlated with mechanical faults, but are susceptible to environmental noise interference. Conventional techniques use the original time domain data or simple fast Fourier transform for frequency domain transformation, which cannot adaptively extract the key frequency bands related to the fault, resulting in insufficient sensitivity to subtle fault modes such as early cracks or bearing wear, and key fault features are easily submerged by noise.

[0071] To address the aforementioned technical issues, embodiments of this application employ a fusion method of adaptive bandpass filtering and wavelet packet decomposition to enhance the temporal data of sample features.

[0072] In one implementation, the sample vibration time series data includes multiple sample acceleration composite amplitudes that correspond one-to-one with multiple sampling times; S130 includes: steps (4) to (7), as detailed below:

[0073] Step (4): Based on the amplitude of the sample vibration time series data at the sample acceleration synthesis amplitude at different values ​​in each sample feature time series data, perform clustering processing on the sample vibration time series data to obtain the number of sample frequency bands corresponding to the normalized vibration sub-vectors of each sample.

[0074] Specifically, in this embodiment, the sample feature time-series data needs to be divided into multiple sub-band sets based on the number of sample frequency bands and the synthesized amplitude of sample acceleration, so as to further determine the energy-significant sub-bands; wherein, the formula for determining the number of sample frequency bands is as follows:

[0075] ;

[0076] In the formula, This indicates the number of sample frequency bands, or the number of sample sub-bands. One; in actual operation, the number of sample frequency bands Through cluster optimization, it was determined that, in practical applications... The value can be 8;

[0077] This represents the composite acceleration amplitude of multiple samples included in the time series data of sample features. The th frequency point The composite amplitude of sample acceleration at each frequency; This represents the number of frequency points corresponding to the time series data of the sample features, i.e., the number of frequency points. indivual; ;

[0078] express In the nth sample sub-band The set of frequencies of each sub-band; express In the nth sample sub-band The centroid frequency of each sub-band; Represents the L2 norm;

[0079] This represents the formula used to minimize the objective function "number of sample bandwidths". value.

[0080] Step (5): Based on the number of sample frequency bands, perform wavelet packet decomposition on the normalized vibration sub-vectors of each sample to obtain the sample wavelet packet frequency band coefficient vectors of each sub-frequency band.

[0081] The sample wavelet packet frequency band coefficient vector includes sample wavelet packet coefficient vectors that correspond one-to-one with the multiple vibration frequencies included in the sub-band.

[0082] Specifically, in the embodiments of this application, the sample normalized vibration time series data subvectors are... Perform wavelet packet decomposition, specifically using the Daubechies wavelet basis, defining the first... The sample wavelet packet frequency band coefficient vector of each sub-band is The wavelet packet coefficient vectors of each sub-band are obtained based on the wavelet packet decomposition function, thereby extracting the features of each sub-band.

[0083] Step (6): Determine the adaptive weights of the samples corresponding to each sub-band based on the wavelet packet coefficient vectors of each sample.

[0084] Specifically, in this embodiment, the total energy of each sub-band is calculated, and then adaptive weights are obtained based on the elements of the sub-band wavelet packet coefficient vector, thereby generating weights for strengthening fault-sensitive frequency bands; wherein, the formula for determining the adaptive weights of the samples is as follows:

[0085] ; ;

[0086] In the formula, Indicates the first Adaptive weighting of samples in each sub-band;

[0087] Indicates the first The total energy of each sub-band; Indicates the first The total energy of each sub-band; the first The sub-band is the same as the first Sub-bands that are different from each other;

[0088] Indicates the first Sample wavelet packet frequency band coefficient vector of each sub-band , including The first sample wavelet packet coefficient vector Wavelet packet coefficients; .

[0089] Step (7): Determine the sample vibration enhancement features based on the sample wavelet packet frequency band coefficient vector and sample adaptive weights corresponding to each sub-band.

[0090] Specifically, in this embodiment, the dimension-reduced sample vibration enhancement features are obtained based on the sample wavelet packet frequency band coefficient vector and the first dimension-reduced projection matrix; wherein, the sample vibration enhancement features are defined. ;

[0091] In the formula, Let represent the first dimension-reduced projection matrix. These are the model parameters trained during the training of the fault mode recognition model. During the training of the fault mode recognition model, the parameters are generated by initialization using a random Gaussian distribution. The elements are sampled from a Gaussian distribution with a mean of 0 and a standard deviation of 0.01. Backpropagation is used to optimize the model during training to reduce feature redundancy and computational complexity while retaining the main fault information.

[0092] Indicates the first The sample wavelet packet frequency band coefficient vector of each sub-band. This represents the sample wavelet packet frequency band coefficient vector of the first sub-band. This represents the sample wavelet packet frequency band coefficient vector of the second sub-band. Indicates the first The sample wavelet packet frequency band coefficient vector of each sub-band; Indicates the first Individual frequency band sample adaptive weights Indicates the first Sub-band sample adaptive weights, This represents the adaptive weights of the samples in the second sub-band. Indicates the first Adaptive weights for samples in each sub-band.

[0093] It should be noted that conventional wavelet packet decomposition treats all sub-bands equally, but the bearing crack fault in the green ammonia reactor will produce a narrowband resonance concentrated in the 2~4kHz range. The term assigns higher weights to fault-sensitive frequency bands through energy proportion weighting. When using the Daubechies wavelet basis, the energy concentration of the fault frequency band exhibits a superlinear relationship with the weights, i.e. The nonlinear amplification effect enhances the characteristic response intensity of early cracks, which cannot be detected by conventional equal-weighting methods or inspired by conventional techniques. More importantly, the sample adaptive weighting... Calculate the first dimension-reduced projection matrix After cascading optimization and weight enhancement of high-frequency bands, the first dimension-reduced projection matrix is ​​formed. During training, it automatically learns the sparse projection direction of the fault frequency band, which enables self-suppression of the noise frequency band and can still extract weak periodic impact components under strong background noise.

[0094] For example, a comparative analysis of time-frequency feature enhancement is presented to demonstrate the differences between conventional Fast Fourier Transform (FFT) and the vibration feature enhancement method provided in the embodiments of this application in vibration signal processing. Figure 4 As shown, Figure 4 This is a comparison chart showing the feature enhancement effect of the vibration time series data provided in the embodiments of this application. Figure 4 The horizontal axis represents frequency (Hz). Figure 4 The vertical axis represents the normalized energy (dimensionless).

[0095] Figure 4 This paper compares the conventional FFT (Fast Fourier Transform) method with the time-frequency enhancement technique of this application (the vibration feature enhancement method provided in the embodiments of this application). The conventional FFT method ( Figure 4 The spectral energy distribution of the dotted line above is uniform but fluctuates violently. There are no significant characteristic peaks in the critical fault frequency bands (170-190 Hz bearing wear sensitive band, 400-440 Hz rotor imbalance sensitive band), indicating that high-frequency noise overwhelms the fault characteristics. The vibration characteristic enhancement method provided in this application embodiment ( Figure 4 The solid line below shows sharp energy peaks in the fault-sensitive frequency band, such as higher amplitude at 180 Hz and 420 Hz. This indicates that after wavelet packet decomposition, the fault-sensitive frequency band is given higher weight by energy ratio. Combined with the noise suppression function of the dimension reduction projection matrix, the narrowband resonance of early bearing cracks is nonlinearly amplified.

[0096] For another example, a three-dimensional spectral waterfall plot is used for frequency band analysis of vibration feature enhancement. The three-dimensional spectral waterfall plot intuitively presents the processing effects of different vibration feature extraction methods in the time and frequency domain, as shown in Figures 5(a), 5(b), and 5(c). Figure 5(a) is a frequency band analysis plot of sample vibration time series data provided in the embodiment of this application. Figure 5(b) is a frequency band analysis plot of traditional wavelet packet decomposition provided in the embodiment of this application. Figure 5(c) is a frequency band analysis plot of vibration feature enhancement method provided in the embodiment of this 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).

[0097] 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 in this application embodiment when processing the same segment 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 submerged by background noise; Figure 5(b) shows a relatively uniform energy distribution, but the fault-sensitive frequency band is not enhanced; Figure 5(c) exhibits obvious feature selectivity, with significant energy enhancement in the 150-250 Hz fault-sensitive frequency band (bearing crack characteristic frequency band), forming a prominent "energy ridge," while the energy in the 400-500 Hz noise frequency band is effectively suppressed. These time-frequency characteristics demonstrate that the vibration feature enhancement method provided in this application embodiment can achieve the dual effects of "fault-sensitive frequency band enhancement" and "noise frequency band suppression," solving the problem that conventional methods cannot adaptively extract key frequency bands, and providing a reliable feature basis for the early diagnosis of mechanical faults.

[0098] S140: Perform feature enhancement processing on the normalized temperature sub-vectors corresponding to the time series data of each sample feature to obtain temperature vibration enhancement features.

[0099] Specifically, in the embodiments of this application, although the temperature time series data is smooth, the fault modes such as abnormal heat distribution in the reactor are manifested as local gradient changes. Conventional techniques use global averaging or sliding window processing, ignoring the spatial dependence between temperature data points, and cannot 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 abrupt changes in smooth data.

[0100] To address the aforementioned technical problems, embodiments of this application employ a local polynomial fitting method based on gradient magnitude weighting to determine the temperature vibration enhancement characteristics.

[0101] In one implementation, the sample temperature time series data includes multiple sample temperature values ​​that correspond one-to-one with multiple sampling times; the sample normalized temperature subvector includes multiple sample temperature vectors that correspond one-to-one with multiple sampling times and multiple sample temperature values; S140 includes steps (8) to (12), as detailed below:

[0102] Step (8): Determine multiple sample temperature windows based on the sample temperature time series data in the feature time series data of each sample.

[0103] The temperature window set includes multiple sample temperature values ​​that correspond one-to-one with multiple consecutive sampling times.

[0104] Specifically, in this embodiment, a set of temperature windows is obtained based on the window size and window sliding step size, thereby establishing a local analysis unit; wherein, the formula for determining the number of sample temperature windows is as follows:

[0105] ;

[0106] In the formula, Indicates the number of sample temperature windows; The dimension representing the time-series temperature data of the sample; This indicates the window size of the sample temperature window. Based on the fault characteristic scale adjustment, it can typically be set between 32 and 128 to ensure that the sample temperature window covers the local temperature gradient. In practice, it can be... The value is set to 64; This indicates a round-down operation; S represents the window sliding step size, used to control the window overlap rate. In practice, the value of S can be set to 32.

[0107] In the set of temperature windows, the first The set of temperature windows is represented as ; ;

[0108] ;

[0109] In the formula, This indicates the temporal position of the sample temperature values ​​included in the temperature window set within the sample temperature time series data. For example, if The value of 5 indicates that the corresponding sample temperature value is the 5th position in the sample temperature time series data;

[0110] Indicates the first The sample temperature value at the first time position among the multiple sample temperature values ​​included in the temperature window set; Indicates the first -1 The sample temperature value at the first time position among the multiple sample temperature values ​​included in the temperature window set; Indicates the first The sample temperature value at the last time position among the multiple sample temperature values ​​included in the temperature window set; This indicates the window sliding step size, used to control the window overlap rate. In actual operation, it can be... The value is set to 32.

[0111] Step (9): Based on the multiple sample temperature vectors included in each sample normalized temperature sub-vector, determine the difference between the sample temperature vectors corresponding to any two adjacent sampling times to obtain multiple sample temperature gradient vectors that correspond one-to-one with the multiple sample temperature vectors.

[0112] Specifically, in this embodiment, gradient values ​​are obtained based on the values ​​of adjacent normalized temperature points to identify potential fault regions; wherein, the formula for determining the sample temperature gradient vector is as follows:

[0113] ;

[0114] In the formula, For the first The sample temperature gradient vector corresponding to each sample temperature vector;

[0115] Represents the first normalized temperature subvector in the sample. A sample temperature vector; Represents the first normalized temperature subvector in the sample. -1 sample temperature vector.

[0116] Sample temperature gradient vector The use of backward difference instead of central difference is determined based on the temperature field characteristics of the green ammonia reactor. The spatial arrangement density of the temperature sensors on the reactor tube wall is 5 cm / point. The temperature gradient change caused by the leakage fault only crosses 2 to 3 sensors in the early stage. Central difference will blur these micro-scale abrupt boundary due to the smoothing effect, while backward difference can amplify the gradient extreme points at the fault edge while maintaining computational efficiency.

[0117] Step (10): Determine the sample gradient weights corresponding to each sample temperature gradient vector based on the sample temperature gradient vector and the preset sensitivity parameters.

[0118] Specifically, in this embodiment, gradient weights are obtained based on gradient values ​​and sensitivity parameters to enhance high gradient regions; wherein, the formula for determining sample gradient weights is as follows:

[0119] ;

[0120] In the formula, Indicates the first The gradient weights corresponding to each sample temperature vector;

[0121] This represents a sensitivity parameter used to control the weight decay rate. In practice, it can be... The value is set to 0.5; This represents the natural exponential function.

[0122] It should be noted that conventional methods often directly take the absolute value or square of the gradient as the weight. However, the normal heat distribution of a green ammonia reactor has a background gradient of ±0.3℃ / cm, while the fault gradient typically ranges from 1 to 3℃ / cm. Therefore, the embodiments of this application use gradient weighting... During the calculation process, the gradient magnitude and weight decay are asymmetric, forming a weight separation effect in the fault zone / background zone. The nonlinear decay automatically weakens the fitting weight in the flat region in the polynomial fitting, forcing the regression curve to closely follow the high gradient region.

[0123] 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 minimizing the objective function.

[0124] Specifically, in this embodiment, polynomial coefficients are obtained based on normalized temperature points within a window and gradient weights to highlight the fault indication area; wherein, the objective function for highlighting the fault indication area is as follows:

[0125] - ;

[0126] In the formula, express The order of the polynomial is the nth The polynomial coefficients of order 1 are obtained by solving a weighted least squares problem on the objective function, and represent the temperature polynomial trend of each sample temperature window. ;

[0127] express of Power of; Represents the polynomial coefficients Minimize the objective function.

[0128] It should be noted that conventional polynomial fitting treats all points within the window equally, but fault-related temperature anomalies are often concentrated locally. Therefore, this embodiment uses gradient weights. A fault-oriented adaptive fitting is achieved. In the low-temperature region caused by blockage, high weights drive the regression curve to concave downwards, while in the high-temperature region caused by leakage, it convex upwards. This is achieved when the polynomial order... When set to 2, the coefficients of the second-order polynomial The sign change can directly indicate the fault type, with negative values ​​indicating blockage and positive values ​​indicating leakage, which is something that cannot be achieved by conventional methods.

[0129] Step (12): Determine the sample temperature enhancement features based on the sample polynomial coefficients corresponding to each sample temperature window.

[0130] Specifically, in this embodiment, the dimension-reduced sample temperature enhancement features are obtained based on the polynomial coefficients of all orders of all sample temperature windows and the second dimension-reduced projection matrix; wherein, the sample temperature enhancement features are defined. ;

[0131] In the formula, Indicates the first The first sample temperature window Polynomial coefficients of order 1 Indicates the first The first sample temperature window Polynomial coefficients of order 1 Indicates the first The first sample temperature window Polynomial coefficients of order 1 Indicates the first The first sample temperature window Polynomial coefficients of order 1 Indicates the first The first sample temperature window Coefficients of a polynomial of order 1; Indicates vector transpose;

[0132] Let represent the second dimension-reduced projection matrix. The parameters of the model trained during the training process of the fault mode recognition model; the second dimension-reduced projection matrix. Principal component analysis is used to initialize the samples using polynomial coefficients of all orders across all temperature windows. Orthogonality must be satisfied to preserve maximum variance, dimensionality reduction is achieved to reduce redundancy, and key trend features are extracted.

[0133] It should be noted that weighted least squares and the second dimension-reduced projection matrix... The synergistic effect enables principal component analysis to automatically focus on the characteristic variance of the weighted regions.

[0134] For example, a study analyzes the effectiveness of temperature gradient-weighted feature extraction using heatmaps, comparing the detection capabilities of different temperature feature extraction methods for localized blockage faults. Figure 6 As shown, Figure 6 This is a comparison chart showing the effectiveness of the temperature-enhanced feature extraction method provided in this application for detecting local blockages. Figure 6The dataset includes three subplots. The x-axis of each subplot represents the reactor length (cm), the y-axis represents the reactor circumference (cm), and the color mapping of the subplots represents the intensity of the temperature characteristic (dimensionless relative value). Specifically, (a) shows the performance of the moving average method in detecting low-temperature zones caused by localized blockages; (b) shows the performance of the global polynomial fitting method in detecting low-temperature zones caused by localized blockages; and (c) shows the performance of the gradient-weighted method described in this technique in detecting low-temperature zones caused by localized blockages.

[0135] Figure 6 The performance of three methods—moving average, global polynomial fitting, and gradient weighting (the temperature feature enhancement method provided in this application)—in detecting low-temperature regions caused by local blockages was compared. Experimental results showed that the heatmap of the moving average method exhibited overly smooth characteristics, and the low-temperature features in the 15-25 cm length and 25-35 cm circumference were severely weakened. Although the heatmap of the global polynomial fitting method reflected the overall trend, it failed to capture the local blockage features at all. The heatmap of the temperature feature enhancement method provided in this application clearly showed the feature enhancement effect of the low-temperature fault area, verifying the core value of gradient weight calculation. The asymmetric weight allocation constructed by the exponential decay function gives higher fitting weights to the high gradient area (fault feature area), thereby accurately capturing local temperature abrupt changes that traditional methods cannot detect in polynomial regression.

[0136] S150: Determine the sample dimensionality reduction temperature-vibration fusion feature vector based on the sample vibration enhancement features and sample temperature enhancement features corresponding to the time series data of each sample feature.

[0137] Specifically, the sample vibration enhancement features and sample temperature enhancement features have different dimensions and statistical properties. Directly connecting the two features will lead to input redundancy and dimensionality mismatch. Conventional techniques such as principal component analysis or simple feature concatenation ignore the interaction between the sample vibration enhancement features and 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.

[0138] To address the aforementioned technical issues, this application employs a dual-branch attention fusion mechanism to achieve feature fusion and dimensionality reduction input.

[0139] In one implementation, S150 includes steps (13) to (16), as detailed below:

[0140] 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.

[0141] Specifically, in this embodiment, a sample vibration feature attention weight vector is obtained based on the current vibration feature attention parameter matrix and the sample vibration enhancement features, which is used to enhance the fault-sensitive component; wherein, the formula for determining the sample vibration feature attention weight vector is as follows:

[0142] ;

[0143] in, This represents the attention weight vector for sample vibration features; This represents the current vibration feature attention parameter matrix. These are the model parameters trained during the training process of the fault mode recognition model;

[0144] Indicates the vibration enhancement characteristics of the sample; express Activation function.

[0145] 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.

[0146] Specifically, in this embodiment, a sample temperature feature attention weight vector is obtained based on the current temperature feature attention parameter matrix and the sample temperature enhancement feature, which is used to strengthen the fault-sensitive component; wherein, the formula for determining the sample temperature feature attention weight vector is as follows:

[0147] ;

[0148] in, This represents the attention weight vector for sample temperature features. This represents the attention parameter matrix for the current temperature feature. These are the model parameters trained during the training process of the fault mode recognition model;

[0149] Indicates the temperature enhancement feature of the sample; express Activation function.

[0150] Step (15): Determine the sample temperature and vibration fusion feature vector based on the sample vibration enhancement feature, sample vibration feature attention weight vector, sample temperature enhancement feature, and sample temperature feature attention weight vector corresponding to each sample feature time series data.

[0151] Specifically, in this embodiment, the temperature and vibration fusion feature vector is obtained based on element-wise multiplication and feature vectors, thereby preserving modal interaction information; wherein, the formula for determining the sample temperature and vibration fusion feature vector is as follows:

[0152] ;

[0153] In the formula, This represents the sample temperature and vibration fusion feature vector; Indicates a vector concatenation operation; This represents element-wise multiplication, used to scale feature elements according to their weights.

[0154] It should be noted that the overall temperature rise or fall trend component in the sample temperature enhancement feature is statistically correlated with the broadband energy in the sample vibration enhancement feature, but the fault sensitivity information of local gradients or specific frequency band energies is uncorrelated. This is achieved through the attention weight vector of the sample temperature feature. and sample vibration feature attention weight vector The cross-modal attention filtering allows important feature elements to be scaled to the [0,2] interval, enabling the construction of a fault subspace projection in the feature space. When the bearing wears, the high-frequency energy weight of vibration and the local gradient weight of temperature are simultaneously enhanced, so that the fusion vector forms a joint activation in the corresponding dimension, which cannot be achieved by the conventional splicing method that does not distinguish the importance of features.

[0155] Step (16): Determine the sample dimensionality reduction temperature and vibration fusion feature vector based on the sample temperature and vibration fusion feature vector corresponding to each sample feature time series data and the preset third dimensionality reduction projection matrix.

[0156] Specifically, in this embodiment, the sample reduced-dimensional temperature and vibration fusion feature vector is obtained based on the sample temperature and vibration fusion feature vector and the third reduced-dimensional projection matrix, thereby generating a compact input representation, as follows:

[0157] ;

[0158] In the formula, This represents the dimensionality-reduced temperature and vibration fusion feature vector of the sample; This represents the sample temperature and vibration fusion feature vector; This represents the third-dimensional reduced projection matrix. To train the model parameters during the training process of the fault mode recognition model, principal component analysis is used to initialize them through the sample temperature and vibration fusion feature vector, thereby reducing training complexity and enhancing feature complementarity.

[0159] S160: The fault mode recognition model is iteratively trained by fusing feature vectors of temperature and vibration from multiple sample dimensionality reduction.

[0160] Specifically, the failure modes of green ammonia reactors are complex, with multiple types of failures 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, making it impossible to adaptively learn the abstract representation of complex failure modes, resulting in insufficient ability to identify complex failure modes.

[0161] To address the aforementioned technical issues, this application's embodiments employ parameterized exponential linear units and combine them with inter-layer feature scaling in the forward propagation of deep neural networks, thereby achieving adaptive deep feature learning for deep neural networks.

[0162] In one implementation, the fault mode recognition model includes: a feature extraction module, which includes: multiple sequentially connected network layers; S160 includes: steps (17) to (20), as detailed below:

[0163] Step (17): For each sample's dimension-reduced temperature and vibration fusion feature vector, the sample's dimension-reduced temperature and vibration fusion feature vector is processed sequentially through the multiple network layers remaining after the last network layer in the feature extraction module to obtain the first sample's output feature vector.

[0164] Step (18): The sample output feature vector is processed by the feature extraction module according to the current weight matrix and current bias corresponding to the last network layer in the feature extraction module to obtain the second sample output feature vector.

[0165] Specifically, in this embodiment, 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:

[0166] ;

[0167] In the formula, This indicates the last (number) feature extraction module. The second sample output feature vector of (each) network layer; Indicating the first feature extraction module The first sample output by each network layer outputs a feature vector. For the first layer, The representation represents the dimensionality reduction and temperature-vibration fusion feature vector of the sample. ; Indicating the first feature extraction module The current weight matrix corresponding to each network layer; Indicating the first feature extraction module The current bias, current weight matrix, and current bias of each network layer are model parameters trained during the training process of the fault mode recognition model.

[0168] Step (19): Using the feature extraction module, determine the 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.

[0169] Specifically, in the embodiments of this application, the feature vector is output based on the second sample. Get the first The output of the activation function of the nth network layer is used to enhance nonlinear expressive power; among which, the output of the activation function of the nth network layer is determined. The formulas for the output of the activation function of each network layer are shown below:

[0170] ;

[0171] In the formula, Indicates the first The output of the activation function of each network layer; This represents the parameterized exponential linear unit activation function; Represents element-wise multiplication; The parameter vector represents the parameterized exponential linear unit. The parameter vector represents the model parameters trained during the training process of the fault mode recognition model. Used to control the curvature in the negative region.

[0172] In this embodiment, 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:

[0173] ;

[0174] In the formula, Indicating the first feature extraction module The sample scaling factor for each network layer;

[0175] This indicates that the first sample outputs a feature vector. The dimension; This represents a logarithmic function, with the default base being the natural constant.

[0176] Step (20): Using the feature extraction module, determine the feature vector of the third sample based on the feature vector of the second sample and the sample scaling factor.

[0177] Specifically, in the embodiments of this application, based on the first The output of the activation function of each network layer and the Sample scaling factor for each network layer The third sample output feature vector is obtained, thus completing feature scaling; the formula for determining the third sample output feature vector is shown below:

[0178] ;

[0179] In the formula, Indicating the first feature extraction module The third sample output feature vector of each network layer.

[0180] It should be noted that while normalization in conventional network layers forces zero mean and unit variance, fault features may exhibit pattern-specific scale distributions in the feature extraction module, such as leakage features having a norm greater than occlusion features, and sample scaling factors. To address the feature scale drift problem in the feature extraction module, a gentle autoscaling mechanism is constructed using a logarithmic function, which works when the feature norm exceeds a certain threshold. hour, Provides sublinear growth, such as hour, It can overcome the feature explosion scenario that is prone to occur in residual networks. The growth rate approaches This forms a direct stabilizing valve.

[0181] In one implementation, the fault mode recognition model further includes a feature classification module; after step (20), the method further includes steps (21) to (24), as detailed below:

[0182] Step (21): Using the feature classification module, determine the prediction probability of the first sample corresponding to each fault mode category based on the current deep feature classification weight vector and the third sample output feature vector corresponding to each fault mode category.

[0183] Specifically, in the embodiments of this application, the feature extraction module in the first... The third sample output feature vector of each network layer contains high-level abstract features, but may weaken local sensitive patterns in the original sensor data, such as tiny temperature abrupt changes. Conventional Softmax classification layers directly use the third sample output feature vector for linear mapping, ignoring the sample reconstructed vector. The lack of detailed information leads to insufficient response to early, minor faults such as micro-leakage.

[0184] To address the aforementioned technical issues, this application employs a residual feature bridging and gating confidence fusion mechanism, which integrates deep abstract features with shallow detailed information to achieve feature classification.

[0185] In this embodiment, the prediction probability of the first sample is obtained based on the output feature vector of the third sample and the current deep feature classification weight vector; wherein, the formula for determining the prediction probability of the first sample is as follows:

[0186] ;

[0187] In the formula, Indicates corresponding to The first of the fault mode categories The first sample predicted probability for each failure mode category; Indicating the first feature extraction module The third sample outputs the feature vector of each network layer;

[0188] Indicates the corresponding to the first The current deep feature classification weight vector for each fault mode category; Indicates the corresponding to the first The current deep feature classification weight vector for each fault mode category; The first failure mode category and the first The fault mode categories are different fault mode categories; the current deep feature classification weight vector is the model parameters trained during the training process of the fault mode recognition model.

[0189] It should be noted that the predicted probability of the first sample In synergy with the residual path, the deep third sample outputs a feature vector. Although it includes high-level abstraction patterns, it weakens the response to local details of early failures by preserving The complete dimension, rather than the dimensionality reduction before the conventional Softmax, enables subsequent residual gating to accurately locate the details lost in the deep network. When the deep feature response of micro-leakage faults is weak, the residual path can extract the spur features in the normalized samples. The two complement each other and significantly improve the early fault recall rate.

[0190] Step (22): Using the feature classification module, determine the sample residual feature vector based on the sample normalized temperature sub-vector, the sample normalized vibration sub-vector, and the current residual projection matrix.

[0191] Specifically, in the embodiments of this application, the sample recombination vector is extracted through the residual path. The fault-sensitive details are identified to determine the sample residual feature vector, thereby extracting the fault-sensitive details. The formula for determining the sample residual feature vector is shown below:

[0192] ;

[0193] In the formula, Represents the sample residual feature vector; Represents the ReLU activation function; This represents the current residual projection matrix, which consists of the model parameters trained during the training of the fault mode recognition model. This represents the sample recombination vector, with a dimension of 4096.

[0194] 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.

[0195] Specifically, in this embodiment, a sample gating weight vector is obtained based on the sample residual feature vector and the current gating parameter matrix, thereby filtering key features; wherein, the formula for determining the sample gating weight vector is as follows:

[0196] ;

[0197] In the formula, Indicates the corresponding to the first Sample gating weight vectors for each failure mode category;

[0198] Indicates the corresponding to the first The current gating parameter matrix for each fault mode category, where the current gating parameter matrix consists of the model parameters trained during the training of the fault mode recognition model.

[0199] Step (24): Using the feature classification module, 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.

[0200] Specifically, in the embodiments of this application, based on the corresponding to the first The sample gating weight vector of the fault mode category and the current residual classification weight vector are used to obtain the corresponding fault mode category. The second sample predicted probability of each fault mode category is used to fuse the gating mechanism; wherein, the probability corresponding to the fault mode category is determined. The formula for the second sample prediction probability of each failure mode category is as follows:

[0201] ;

[0202] In the formula, This indicates that the value corresponding to the first residual feature is based on the residual feature. The second sample predicted probability for each failure mode category; Indicates the corresponding to the first Sample gating weight vectors for each failure mode category; Indicates the first The current residual classification weight vector of the class; Indicates the first The current residual classification weight vector of the class, which is the model parameters trained during the training of the fault mode recognition model.

[0203] It should be noted that conventional residual connections directly fuse features, while the embodiments in this application learn independent sample gating weight vectors for each type of fault mode. For example, for leakage faults, the sample gating weight vector automatically strengthens the dimension corresponding to the gradient 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 fault features at the same time, the sample gating weight vector makes the probability calculation of each type only focus on its sensitive dimension, avoiding misjudgment caused by feature mutual exclusion.

[0204] In one implementation, the fault mode recognition model further includes a feature classification module; after step (24), the method further includes step (25), as detailed below:

[0205] Step (25): Using the feature classification module, determine the third sample prediction probability corresponding to each fault mode category based on the first sample prediction probability and the second sample prediction probability corresponding to each fault mode category.

[0206] The third sample prediction probability indicates the failure mode category of the green ammonia reactor.

[0207] Specifically, in this embodiment, 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 fusing deep abstraction and shallow detail information; wherein, the formula for determining the third sample prediction probability is as follows:

[0208] ;

[0209] ;

[0210] ;

[0211] In the formula, Indicates the probability predicted based on the first sample. The determined corresponding to the first Confidence level of each failure mode category; Indicates the probability predicted based on the first sample. The determined corresponding to the first Confidence level of each failure mode category; Indicates the probability predicted based on the second sample. The determined corresponding to the first Confidence level of each failure mode category; Indicates the probability predicted based on the second sample. The determined corresponding to the first Confidence level of each failure mode category; Indicates the corresponding to the first The third sample predicted probability for each failure mode category.

[0212] In one implementation, after step (25), the method further includes steps (26) to (29), as detailed below:

[0213] Step (26): Determine the sample frequency weight corresponding to each fault mode category based on the number of samples in the time series data of multiple sample features corresponding to each fault mode category.

[0214] Specifically, the sample distribution of fault mode categories such as normal, leakage, and blockage is often unbalanced. Conventional cross-entropy loss functions ignore the weighting issues caused by the difference in the number of samples of different categories, resulting in a low recognition rate for a few categories such as severe faults. This makes it impossible to effectively handle the dual challenges of differences in category frequency and classification difficulty.

[0215] To address the aforementioned technical issues, this application employs a weighted focus loss function, integrating sample frequency weights, sample focus modulation terms, labels, and third-sample prediction probabilities to enhance the model's sensitivity to rare faults. Before calculating the loss function value using the weighted focus loss function, it is necessary to first calculate the sample frequency weights corresponding to each fault mode category. The formula for determining the sample frequency weights is as follows:

[0216] ;

[0217] In the formula, This indicates the quantity of time-series data representing sample features;

[0218] express The label of the time series data of the sample features is the first. The number of time-series data of sample features for each fault mode category; ; Indicates the number of failure mode categories;

[0219] Indicates the corresponding to the first Sample frequency weights for each failure mode category.

[0220] Step (27): Determine the sample difficulty factor corresponding to each fault mode category based on the sample frequency weight and historical classification accuracy of each fault mode category.

[0221] Specifically, in this embodiment, a sample difficulty factor is determined based on sample frequency weights and historical classification accuracy to reflect the classification difficulty; the formula for determining the sample difficulty factor is as follows:

[0222] - ;

[0223] In the formula, Indicates the corresponding to the first The sample difficulty factor for each failure mode category;

[0224] Indicates the corresponding to the first Historical classification accuracy of each failure mode category This was determined based on the previous training round; This indicates an adjustable parameter used to control the difficulty sensitivity. In actual operation, it can be... The value is set to 2.

[0225] Step (28): Determine the sample focus modulation term corresponding to each fault mode category based on the third sample prediction probability and sample difficulty factor corresponding to each fault mode category.

[0226] Specifically, in this embodiment, the 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:

[0227] ;

[0228] In the formula, Indicates the corresponding to the first Sample focus modulation terms for each fault mode category;

[0229] Indicates the corresponding to the first The third sample predicted probability for each failure mode category.

[0230] Step (29): Determine the loss function value based on the third sample prediction probability, sample frequency weight, and sample focus modulation term corresponding to each fault mode category.

[0231] The loss function value is used to update the model parameters of the fault mode recognition model.

[0232] Specifically, in this embodiment, the loss function value is obtained by using sample frequency weights, sample focus modulation terms, labels, and the prediction probability of the third sample, thereby enhancing the sensitivity to rare faults; the formula for calculating the loss function value is as follows:

[0233] ;

[0234] In the formula, This represents the loss function value calculated based on the weighted focus loss function; Indicates the corresponding to the first One-hot encoding of the labels for each failure mode category.

[0235] It should be noted that the sample frequency weighting Directly combating sample imbalance by increasing the weight of rare fault loss and the sample difficulty factor. The system dynamically senses the classification difficulty and addresses categories with historically low accuracy, such as the confusion between blockage and scaling. Increase the sample focus modulation term More aggressive suppression of the loss contribution of easily separable samples and sample focus modulation term The exponential form focuses on difficult samples, when hour The loss of easily distinguishable samples decays, while hour Difficult sample loss is preserved; the coupling of the triple regulation generates adaptive learning for fault difficulty; faults are easily identified with a large sample size; and the loss is... and Dual inhibition, for faults that are difficult to distinguish with low sample size, is... Magnified and maintain.

[0236] In this embodiment, during the iterative training of the fault mode recognition model, a mini-batch gradient descent strategy is used for iterative optimization. In each iteration, 256 samples of time-series feature data are randomly sampled from the training set to form a batch of data. Forward propagation is performed to calculate the predicted probability, the loss function value is calculated based on the weighted focus loss function, and the gradient of the loss function with respect to all trainable parameters of the model is calculated through the backpropagation algorithm. The parameters are updated using an adaptive moment estimation optimizer. After each training cycle, a weighted F1 score is calculated on the validation set as a monitoring indicator for the early stopping strategy.

[0237] In this embodiment, the training termination conditions for the fault mode recognition model include two mechanisms. The first is early stopping based on performance convergence, which means that when the weighted F1 score on the validation set does not improve for 10 consecutive training cycles (improvement threshold < 0.001), the current optimal model parameters are saved and training is terminated. The second is forced stopping based on the maximum number of iterations, which means that the preset training cycle limit is 200 cycles. If early stopping is not triggered when the cycle limit is reached, the parameters corresponding to the cycle with the highest F1 score on the validation set are selected as the final model.

[0238] In practice, the loss curve and validation metrics are recorded every cycle during training. When the loss function value is NaN (Not a Number, a special symbol for "invalid value") or the validation set performance drops by more than 20%, it indicates that the model has crashed. Training is immediately interrupted and rolled back to the previous cycle to reinitialize the optimizer state.

[0239] Distance explanation: A comparative analysis of fault identification accuracy is now conducted to evaluate the overall performance of the fault identification system. Figure 7 As shown, Figure 7 A comparison diagram of different fault mode identification methods provided in the embodiments of this application. Figure 7 The horizontal axis represents the labels corresponding to the seven fault identification categories. Figure 7 The vertical axis represents the classification accuracy (dimensionless). Figure 7 The recognition performance of SVM (Support Vector Machine), Random Forest, conventional CNN (Convolutional Neural Network) and the present technology (the green ammonia reactor fault mode recognition method provided in the embodiments of this application) was compared.

[0240] according to Figure 7 It is known that conventional methods such as SVM, random forest, and conventional CNN perform the weakest in "rotor imbalance" and "bearing wear", with a classification accuracy of less than 0.85, and fail to handle the heterogeneity of multimodal features. The embodiments of this application have the highest classification accuracy in all fault mode categories, especially significantly improving "rotor imbalance" and "bearing wear". Experimental results show that: 1) grouping normalization preserves the complementarity of cross-modal features; 2) time-frequency / temperature feature enhancement mechanism extracts weak fault signals; 3) dual-branch attention fusion enables temperature gradient weights and vibration high-frequency weights to be activated in synergy.

[0241] S220: For the target vibration time series data and the target temperature time series data, feature enhancement processing and feature fusion processing are performed sequentially to obtain the target dimension-reduced temperature and vibration fusion feature vector.

[0242] Specifically, in the embodiments of this application, the specific implementation details of S220 can be referred to steps (13) to (16), etc., and will not be repeated here.

[0243] S230: Based on the target's reduced-dimensional temperature and vibration fusion feature vector, the fault mode category corresponding to the target feature time series data is determined through a pre-trained fault mode recognition model.

[0244] Specifically, in the embodiments of this application, when the target dimensionality-reduced temperature and vibration fusion feature vector is input into the 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.

[0245] Second, this application provides a fault mode identification device for a green ammonia reactor, such as... Figure 8 As shown, Figure 8 This is a structural diagram of the green ammonia reactor fault mode identification device provided in an embodiment of this application. The device includes: a data acquisition module 310, a data processing module 320, and a prediction module 330.

[0246] The data acquisition module 310 is used to acquire the target characteristic time series data of the target green ammonia reactor within the target time period; wherein, the target characteristic time series data includes: target vibration time series data and target temperature time series data;

[0247] 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 the target dimension-reduced temperature and vibration fusion feature vector;

[0248] The prediction module 330 is used to determine the fault mode category corresponding to the target feature time series data based on the target dimensionality-reduced temperature and vibration fusion feature vector through a pre-trained fault mode recognition model.

[0249] In one implementation, the apparatus includes: a training module;

[0250] The training module is used to acquire time-series data of multiple sample features from the green ammonia reactor within the sample time period; among which, the sample feature time-series data includes: sample vibration time-series data and sample temperature time-series data;

[0251] The training module is also used to normalize the sample temperature time series data and 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 temperature time series data of the green ammonia reactor, respectively, to obtain the sample normalized temperature sub-vector and sample normalized vibration sub-vector.

[0252] The training module is also used to perform feature enhancement processing on the normalized vibration subvectors corresponding to each sample feature time series data to obtain sample vibration enhancement features.

[0253] The training module is also used to perform feature enhancement processing on the normalized temperature subvectors corresponding to the time series data of each sample feature to obtain sample temperature enhancement features.

[0254] The training module is also used to determine the sample dimensionality reduction temperature and vibration fusion feature vector based on the sample vibration enhancement features and sample temperature enhancement features corresponding to the time series data of each sample feature.

[0255] The training module is also used to iteratively train the fault mode recognition model by using the dimensionality-reduced temperature and vibration fusion feature vectors from multiple samples.

[0256] In one implementation, the sample vibration time series data includes multiple sample acceleration composite amplitudes that correspond one-to-one with multiple sampling times; the training module is also used to perform clustering processing on the sample vibration time series data based on the amplitudes of the sample vibration time series data at sample acceleration composite amplitudes with different values ​​in each 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.

[0257] The training module is also used to perform wavelet packet decomposition on the normalized vibration sub-vectors of each sample according to the number of sample frequency bands to obtain the sample wavelet packet frequency band coefficient vectors of each sub-frequency band.

[0258] The sample wavelet packet frequency band coefficient vector includes: sample wavelet packet coefficient vectors that correspond one-to-one with the multiple vibration frequencies included in the sub-band;

[0259] The training module is also used to determine the adaptive weights of the samples corresponding to each sub-band based on the wavelet packet coefficient vectors of the multiple samples corresponding to each sub-band.

[0260] The training module is also used to determine the vibration enhancement features of the samples based on the wavelet packet frequency band coefficient vector and the sample adaptive weights corresponding to each sub-band.

[0261] In one implementation, the sample temperature time series data includes multiple sample temperature values ​​that correspond one-to-one with multiple sampling times; the sample normalized temperature subvector includes multiple sample temperature vectors that correspond one-to-one with multiple sampling times and multiple sample temperature values; the training module is also used to determine multiple sample temperature windows based on the sample temperature time series data in each sample feature time series data; wherein, the temperature window set includes: multiple sample temperature values ​​that correspond one-to-one with multiple consecutive sampling times.

[0262] The training module is also used to determine the difference between the sample temperature vectors corresponding to any two adjacent sampling times based on the multiple sample temperature vectors included in each sample normalized temperature subvector, and to obtain multiple sample temperature gradient vectors that are in one-to-one correspondence with the multiple sample temperature vectors.

[0263] The training module is also used to determine the sample gradient weights corresponding to each sample temperature gradient vector based on the sample temperature gradient vector and the preset sensitivity parameters.

[0264] The training module is also used to determine the sample polynomial coefficients corresponding to each sample temperature window under the constraint of minimizing the objective function, based on the multiple sample temperature vectors, multiple sample gradient weights and the minimizing objective function corresponding to each sample temperature window.

[0265] The training module is also used to determine the sample temperature enhancement features based on the sample polynomial coefficients corresponding to each sample temperature window.

[0266] In one implementation, the training module is further used to 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.

[0267] The training module is also used to 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.

[0268] The training module is also used to determine the sample temperature and vibration fusion feature vector based on the sample vibration enhancement features, sample vibration feature attention weight vector, sample temperature enhancement features, and sample temperature feature attention weight vector corresponding to the time series data of each sample feature.

[0269] The training module is also used to determine the sample dimensionality reduction temperature and vibration fusion feature vector based on the sample temperature and vibration fusion feature vector corresponding to each sample feature time series data and the preset third dimensionality reduction projection matrix.

[0270] In one implementation, the fault mode recognition model includes: a feature extraction module, which includes: multiple sequentially connected network layers; and a training module, which is further used to process the temperature and vibration fusion feature vector of each sample in turn through the multiple network layers remaining after removing the last network layer in the feature extraction module to obtain the output feature vector of the first sample.

[0271] The training module is also used to process the sample output feature vector through the feature extraction module, based on the current weight matrix and current bias corresponding to the last network layer in the feature extraction module, to obtain the second sample output feature vector;

[0272] The training module is also used to determine the 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, through the feature extraction module.

[0273] The training module is also used to determine the output feature vector of the third sample based on the output feature vector of the second sample and the sample scaling factor through the feature extraction module.

[0274] In one implementation, the fault mode recognition model further includes: a feature classification module; and a training module, which is further used to determine the first sample prediction probability corresponding to each fault mode category by means of the feature classification module, based on the current depth feature classification weight vector corresponding to each fault mode category and the third sample output feature vector.

[0275] The training module is also used to determine the sample residual feature vector by means of the feature classification module, based on the sample normalized temperature sub-vector, the sample normalized vibration sub-vector and the current residual projection matrix;

[0276] The training module is also used to determine the sample gating weight vector based on the sample residual feature vector and the current gating parameter matrix through the feature classification module;

[0277] The training module is also used to determine the second sample prediction probability corresponding to each fault mode category by means of the feature classification module, based on the current gating parameter matrix, the sample residual feature vector corresponding to each fault mode category, and the sample gating weight vector.

[0278] In one implementation, the training module is further configured to determine, through the feature classification module, a third sample prediction probability corresponding to each fault mode category based on the first sample prediction probability and the second sample prediction probability corresponding to each fault mode category.

[0279] The third sample prediction probability indicates the failure mode category of the green ammonia reactor.

[0280] In one implementation, the training module is also used to determine the sample frequency weights corresponding to each fault mode category based on the number of samples corresponding to each fault mode category in the time series data of multiple sample features.

[0281] The training module is also used to determine the sample difficulty factor corresponding to each fault mode category based on the sample frequency weight and historical classification accuracy of each fault mode category.

[0282] The training module is also used to determine the sample focus modulation term corresponding to each fault mode category based on the third sample prediction probability and sample difficulty factor corresponding to each fault mode category.

[0283] The training module is also used to determine the loss function value based on the third sample prediction probability, sample frequency weight, and sample focus modulation term corresponding to each fault mode category.

[0284] The loss function value is used to update the model parameters of the fault mode recognition model.

[0285] Third, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps S210 to S230 provided in the above embodiments.

[0286] Fourth, this application also provides a computer-readable storage medium storing a computer program, wherein the computer program is executed by a processor to perform the steps of S210 to S230 of the above embodiments.

[0287] 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 methods in the preceding method embodiments. For specific implementation, please refer to the steps of S210 to S230 of the method embodiments, which will not be repeated here.

[0288] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0289] Furthermore, the units described as separate components may or may not be physically separate. 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 can be selected to achieve the purpose of this embodiment according to actual needs.

[0290] Furthermore, the functional modules in the various embodiments of this 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.

[0291] It should be noted that if the function is implemented as a software functional 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 this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0292] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.

[0293] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for identifying failure modes in a green ammonia reactor, characterized in that, The method includes: Acquire 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; For the target vibration time series data and the target temperature time series data, feature enhancement processing and feature fusion processing are performed sequentially to obtain the target dimension-reduced temperature and vibration fusion feature vector; Based on the target reduced-dimensional temperature and vibration fusion feature vector, the fault mode category corresponding to the target feature time series data is determined by a pre-trained fault mode recognition model; Before acquiring the target characteristic time-series data of the target green ammonia reactor within the target time period, the method further includes: acquiring multiple sample characteristic time-series data of the sample green ammonia reactor within the sample time period; wherein, the sample characteristic time-series data includes: sample vibration time-series data and sample temperature time-series data; the sample vibration time-series data includes multiple sample acceleration composite amplitudes that correspond one-to-one with multiple sampling times; the sample temperature time-series data includes multiple sample temperature values ​​that correspond one-to-one with multiple sampling times; the sample normalized temperature sub-vector includes multiple sample temperature vectors that correspond one-to-one with multiple sampling times and multiple sample temperature values; Based on the amplitude of the sample vibration time series data at different sample acceleration composite amplitudes in the 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 each sample normalized vibration sub-vector; based on the number of sample frequency bands, wavelet packet decomposition processing is performed on each sample normalized vibration sub-vector to obtain the sample wavelet packet frequency band coefficient vector of each sub-frequency band; wherein, the sample wavelet packet frequency band coefficient vector includes: sample wavelet packet coefficient vectors that correspond one-to-one with multiple vibration frequencies included in the sub-frequency band; based on the multiple sample wavelet packet coefficient vectors corresponding to each sub-frequency band, the sample adaptive weight corresponding to each sub-frequency band is determined; based on the sample wavelet packet frequency band coefficient vectors corresponding to each sub-frequency band and the sample adaptive weight, the sample vibration enhancement feature is determined; Based on the sample temperature time series data in the sample feature time series data of each sample, multiple sample temperature windows are determined; wherein, the temperature window set includes: multiple sample temperature values ​​that correspond one-to-one with multiple consecutive sampling times; based on the multiple sample temperature vectors included in each sample normalized temperature sub-vector, the difference between the sample temperature vectors corresponding to any two adjacent sampling times is determined to obtain multiple sample temperature gradient vectors that correspond one-to-one with multiple sample temperature vectors; based on the sample temperature gradient vectors and a preset sensitivity parameter, sample gradient weights corresponding to each sample temperature gradient vector are determined; based on the multiple sample temperature vectors corresponding to each sample temperature window, the multiple sample gradient weights, and the minimization objective function, the sample polynomial coefficients corresponding to each sample temperature window are determined under the constraint of the minimization objective function; based on the sample polynomial coefficients corresponding to each sample temperature window, sample temperature enhancement features are determined; The fault mode recognition model includes a feature extraction module and a feature classification module. The feature extraction module includes multiple sequentially connected network layers. Iterative training of the fault mode recognition model using multiple sample dimensionality-reduced temperature and vibration fusion feature vectors includes: for each sample dimensionality-reduced temperature and vibration fusion feature vector, processing the sample dimensionality-reduced temperature and vibration fusion feature vector sequentially through the multiple network layers remaining after removing the last network layer in the feature extraction module to obtain a first sample output feature vector; 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 to obtain a second sample output feature vector; determining the sample scaling factor of the last network layer in the feature extraction module based on the first sample output feature vector and its dimension; and determining a third sample output feature vector based on the second sample output feature vector and the sample scaling factor. The third sample output feature vector is used by the feature classification module to determine the fault mode category.

2. The method according to claim 1, characterized in that, Before acquiring the target characteristic time-series data of the target green ammonia reactor within the target time period, the method further includes: Acquire time-series data of multiple sample characteristics from a green ammonia reactor within a sample time period; wherein, the time-series data of sample characteristics includes: sample vibration time-series data and sample temperature time-series data; Based on the noise characteristics corresponding to the vibration time series data and 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 according to the noise characteristics to obtain the sample normalized temperature sub-vector and the sample normalized vibration sub-vector. Feature enhancement processing is performed on the normalized vibration subvectors corresponding to each of the sample feature time series data to obtain sample vibration enhancement features; Feature enhancement processing is performed on the normalized temperature subvector corresponding to each of the sample feature time series data to obtain sample temperature enhancement features. Based on the vibration enhancement feature and temperature enhancement feature of each sample corresponding to the time series data of each sample feature, a sample dimension-reduced temperature-vibration fusion feature vector is determined. The fault mode recognition model is iteratively trained by using multiple sample-dimension-reduced temperature and vibration fusion feature vectors.

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

4. The method according to claim 1, 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: The feature classification module determines the prediction probability of the first sample 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 feature classification module determines the sample residual feature vector based on the sample normalized temperature sub-vector, the sample normalized vibration sub-vector, and the current residual projection matrix. The feature classification module determines the sample gating weight vector based on the sample residual feature vector and the current gating parameter matrix. The feature classification module determines 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.

5. The method according to claim 4, characterized in that, After determining the second sample prediction probability corresponding to each of the fault mode categories, the method further includes: The feature classification module determines the third sample prediction probability corresponding to each of the fault mode categories based on the first sample prediction probability and the second sample prediction probability corresponding to each of the fault mode categories. The third sample prediction probability indicates the failure mode category of the green ammonia reactor.

6. The method according to claim 4, characterized in that, After determining the third sample prediction probability corresponding to each of the fault mode categories, the method further includes: Based on the number of samples corresponding to each of the fault mode categories in the time series data of multiple sample features, the sample frequency weight corresponding to each of the fault mode categories is determined; Based on the sample frequency weight and historical classification accuracy corresponding to each of the fault mode categories, the sample difficulty factor corresponding to each of the fault mode categories is determined. Based on the third sample prediction probability and the sample difficulty factor corresponding to each of the fault mode categories, determine the sample focus modulation term corresponding to each of the fault mode categories; The loss function value is determined based on 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.

7. A fault mode identification device for a green ammonia reactor, characterized in that, The green ammonia reactor failure mode identification device is used to execute the green ammonia reactor failure mode identification method according to any one of claims 1-6, and the device includes: a data acquisition module, a data processing module, and a prediction module; The data acquisition module is used to acquire the 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 the target dimension-reduced temperature and vibration fusion feature vector. The prediction module is used to determine the fault mode category corresponding to the target feature time series data based on the target reduced temperature and vibration fusion feature vector through a pre-trained fault mode recognition model.

Images