Equipment fault identification method and system based on double-domain sensing path

Abstract

The invention relates to the technical field of equipment fault diagnosis, and discloses an equipment fault identification method and system based on a double-domain sensing path, and the method comprises the steps: obtaining an original signal of industrial equipment; constructing a wide area sensing path, adopting an attention-LSTM structure, combining empirical mode decomposition and an improved multi-frequency position coding mechanism, and extracting long-term degradation characteristics of the industrial equipment; a local domain sensing path is constructed, an LSTM-attention structure is adopted, high-order differential and a sparse mask mechanism are combined, and transient fault features of the industrial equipment are extracted; and carrying out dimension alignment on the long-term degradation feature of the wide-area sensing path and the transient fault feature of the local-area sensing path, and outputting a fault identification result of the industrial equipment. According to the invention, through the construction of the double-domain sensing path, the long-term degradation features and the transient fault features are fully fused, and the fault of the industrial equipment can be identified more comprehensively and accurately.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment fault diagnosis, in particular to an equipment fault identification method and system based on a dual-domain perception path. BACKGROUND

[0002] Industrial equipment is prone to performance degradation and even failure during long-term operation due to factors such as load fluctuations, environmental erosion, and component wear. Timely and accurate fault identification is crucial for ensuring the continuity of industrial production and reducing maintenance costs.

[0003] Existing equipment fault identification methods mostly use single-dimensional signal processing methods, which have obvious limitations: methods that rely solely on long-term time series feature extraction cannot capture the sudden characteristics of transient faults; methods that only focus on local transient signals may overlook the cumulative rules in the long-term degradation process of equipment. In addition, traditional methods have problems such as insufficient signal redundancy processing, loss of position information, and mismatched feature dimensions during feature extraction, resulting in low fault identification accuracy and delayed response, which cannot meet the high-precision and real-time fault diagnosis requirements of industrial equipment. Therefore, there is an urgent need for a fault identification scheme that can simultaneously consider long-term degradation rules and transient fault characteristics, and has efficient feature processing and fusion capabilities to improve the reliability and timeliness of industrial equipment fault diagnosis. SUMMARY

[0004] The purpose of the present application is to provide an equipment fault identification method based on a dual-domain perception path, which aims to solve the problem that existing equipment fault identification methods mostly use single-dimensional signal processing methods, which have obvious limitations, and have problems such as insufficient signal redundancy processing, loss of position information, and mismatched feature dimensions during feature extraction.

[0005] The present application provides an equipment fault identification method based on a dual-domain perception path, comprising: Real-time monitoring of industrial equipment to obtain the original signals of the industrial equipment; Constructing a wide-area perception path, using an attention-LSTM structure, combining empirical mode decomposition and an improved multi-frequency position encoding mechanism to process the original signals and extract long-term degradation features of the industrial equipment; Constructing a local domain perception path, using an LSTM-attention structure, combining high-order differentiation and a sparse mask mechanism to process the original signals and extract transient fault features of the industrial equipment; Aligning the long-term degradation features of the wide-area perception path with the transient fault features of the local domain perception path, and outputting the fault identification results of the industrial equipment.

[0006] Preferably, a wide-area perception path is constructed, an attention-LSTM structure is adopted, and an empirical mode decomposition and an improved multi-frequency position encoding mechanism are combined to process the original signal, extract long-term degradation features of the industrial equipment, including: The original signal is subjected to empirical mode decomposition to obtain a plurality of intrinsic mode function components and a residual term. Each intrinsic mode function component is subjected to feature extraction and weight calculation, and the intrinsic mode function components with weight coefficients exceeding a preset threshold are screened for signal reconstruction. The reconstructed intrinsic mode function components are subjected to position encoding through a multi-frequency position encoding mechanism. A multi-head attention mechanism is used for grouping, and the grouping results are input into an LSTM network, and a residual connection is introduced to extract long-term degradation features of the industrial equipment.

[0007] Preferably, when the reconstructed intrinsic mode function components are subjected to position encoding through the multi-frequency position encoding mechanism, the following calculation formula is used: ; Where PEj(k) represents the encoding value of position k in the jth dimension, M represents the total number of reconstructed intrinsic mode function components, t represents the physical change period corresponding to the intrinsic mode function component, j represents the dimension index of position encoding, and n is a positive integer.

[0008] Preferably, when the multi-head attention mechanism is used for grouping, the following calculation formula is used: ; Where Q represents the query matrix, K represents the key matrix, V represents the value matrix, dk represents the dimension of the key vector, and softmax represents the activation function.

[0009] Preferably, a local-area perception path is constructed, an LSTM-attention structure is adopted, and a high-order differential and a sparse mask mechanism are combined to process the original signal, extract instantaneous fault features of the industrial equipment, including: The original signal is subjected to normalization processing, and differential features are extracted through a high-order differential technique, including first-order differential features, second-order differential features, and third-order differential features. The differential features are input into an improved LSTM network, the original embedded features are fused through a gating mechanism, an attention mechanism is constructed for the differential features output by the LSTM network, and a sparse mask corresponding to the differential order is set to extract instantaneous fault features of the industrial equipment.

[0010] Preferably, when the differential features are extracted through the high-order differential technique, the following calculation formula is used: ; Wherein, the first-order differential characteristic is represented by Δ1S(t), the second-order differential characteristic is represented by Δ2S(t), the third-order differential characteristic is represented by Δ3S(t), the normalized original signal at t time is represented by S(t), the normalized original signal at t+1 time is represented by S(t+1), the normalized original signal at t-1 time is represented by S(t-1), and the normalized original signal at t+0 time is represented by S(t+2).

[0011] Preferably, the differential characteristic is input into an improved LSTM network, and when the embedded features are fused through a gating mechanism, the following calculation formula is used: ; Wherein, o represents the output result of the LSTM, f represents the embedded feature, cos(o,f) represents the cosine similarity between the LSTM output feature o and the embedded feature f, g represents the gating coefficient, sigmoid represents the activation function, and G(g1,...,gt) represents the output feature.

[0012] Preferably, a sparse mask corresponding to the differential order is set, including: setting a corresponding time window bandwidth for the differential characteristic; specifically: The first-order differential bandwidth 3z is set for the first-order differential characteristic, the second-order differential bandwidth 2z is set for the second-order differential characteristic, and the third-order differential bandwidth 1z is set for the third-order differential characteristic; wherein, z is a preset constant.

[0013] Preferably, the long-term degradation feature of the wide-area perception path and the instantaneous fault feature of the local-area perception path are dimensionally aligned, including: A hybrid loss function is set, the hybrid loss function includes a wide-area loss function L1 and a local-area loss function L2, and the expression of the hybrid loss function is: L=αL1+ (1-α)L2, wherein, L represents the total loss function, and α represents a dynamic adjustment coefficient; The hybrid loss function is used to realize the collaborative optimization of the wide-area perception path and the local-area perception path.

[0014] The application also discloses a device fault identification system based on a double-domain perception path, which is used for applying the device fault identification method based on the double-domain perception path. The signal acquisition module is configured to monitor the industrial equipment in real time and acquire the original signal of the industrial equipment. The wide-area perception path construction module is configured to construct a wide-area perception path, adopt an attention-LSTM structure, process the original signal by combining an empirical mode decomposition and an improved multi-frequency position encoding mechanism, and extract long-term degradation features of the industrial equipment. The local domain perception path construction module is configured to construct a local domain perception path, adopts an LSTM-attention structure, processes the original signal by combining a high-order differential and a sparse mask mechanism, and extracts instantaneous fault features of the industrial equipment; The fault identification output module is configured to perform dimension alignment on long-term degradation features of the wide-area perception path and instantaneous fault features of the local domain perception path, and output a fault identification result of the industrial equipment.

[0015] Compared with the prior art, the beneficial effects of the present application are that the present application adopts a dual-domain perception path design, the wide-area path captures the long-term degradation law of the equipment, the local domain path captures the instantaneous fault features, multi-dimensional feature complementation is achieved, and the comprehensiveness and accuracy of fault identification are significantly improved. Combined with empirical mode decomposition and high-order differential technology, the long-term and instantaneous feature extraction effects are respectively optimized, and improved multi-frequency position encoding and sparse mask mechanism are used to effectively solve the problems of signal redundancy, position information loss and feature interference. The dual-path collaborative optimization and feature dimension alignment are realized through a hybrid loss function, the rationality and reliability of feature fusion are ensured, and the adaptability of the method to different types of industrial equipment and complex fault scenes is improved. The system module design is simple and efficient, supports real-time signal processing and fault output, can be directly integrated into the existing industrial monitoring system, has low operation and maintenance cost, and has strong practicality. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only part of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the provided drawings.

[0017] Figure 1 is a flowchart of a device fault identification method based on a dual-domain perception path of the present application; Figure 2 is a functional block diagram of a device fault identification system based on a dual-domain perception path of the present application. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be described clearly and completely in the following combined with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0019] This invention provides a method and system for equipment fault identification based on a dual-domain sensing path. By constructing a dual-domain sensing path, it fully integrates long-term degradation features and instantaneous fault features, enabling more comprehensive and accurate identification of faults in industrial equipment. On one hand, the wide-domain sensing path utilizes an attention-LSTM structure, empirical mode decomposition, and an improved multi-frequency position coding mechanism to effectively capture the degradation trend of industrial equipment over a longer time scale. This helps to detect potential fault hazards in advance and provides a strong basis for preventive maintenance. For example, problems such as component wear and performance degradation caused by long-term operation can be detected in a timely manner through the extraction of long-term degradation features. On the other hand, the local-domain sensing path adopts an LSTM-attention structure, combined with high-order differentiation and sparse masking mechanisms, focusing on extracting instantaneous fault features of industrial equipment. When equipment experiences sudden failure, it can quickly and accurately capture the feature information of the fault moment, enabling timely response and diagnosis. For example, when equipment experiences sudden short circuits, overloads, or other instantaneous faults, this path can quickly identify and output fault signals. By aligning these two different types of features dimensionally, the fault identification results are output, avoiding the limitations of single-feature identification and improving the accuracy and reliability of fault identification. This dual-domain sensing fault identification method can significantly reduce equipment downtime, lower maintenance costs, and improve production efficiency.

[0020] like Figure 1 As shown, in one aspect, the present invention discloses a device fault identification method based on a dual-domain sensing path, comprising: S1 performs real-time monitoring of industrial equipment and acquires the raw signals from the industrial equipment.

[0021] In this embodiment, various types of sensors are installed on industrial equipment to collect raw signals.

[0022] Specifically, industrial equipment can be key equipment in multiple fields, such as aircraft engine bearings, wind turbine gearboxes, and energy storage battery systems.

[0023] S2, construct a wide-area perception path, adopt an attention-LSTM structure, and process the original signal by combining empirical mode decomposition and an improved multi-frequency position coding mechanism to extract the long-term degradation characteristics of industrial equipment.

[0024] In some embodiments of the present application, a wide-area perception path is constructed, an attention-LSTM structure is adopted, and an empirical mode decomposition and an improved multi-frequency position encoding mechanism are combined to process the original signal and extract long-term degradation features of industrial equipment, including: performing empirical mode decomposition on the original signal to obtain a plurality of intrinsic mode function components and a residual term; performing feature extraction and weight calculation on each intrinsic mode function component, and reconstructing the signal by screening the intrinsic mode function components with weight coefficients exceeding a preset threshold; performing position encoding on the reconstructed intrinsic mode function components through a multi-frequency position encoding mechanism; grouping by using a multi-head attention mechanism, and inputting the grouping result into an LSTM network to extract long-term degradation features of industrial equipment by introducing a residual connection.

[0025] It can be understood that the empirical mode decomposition can adaptively decompose the original signal into intrinsic mode function components and a residual term, clearly revealing the internal characteristics and variation rules of the signal, and providing a rich and effective data basis for subsequent feature extraction. The intrinsic mode function components with weight coefficients exceeding the preset threshold are screened for signal reconstruction, highlighting the signal components that have an important influence on the long-term degradation characteristics of the equipment, removing the interference of noise and irrelevant information, and making the extracted features more accurate and targeted. The multi-frequency position encoding mechanism gives the reconstructed intrinsic mode function components position information, which helps the model better understand the change relationship of the signal at different time scales and enhances the processing capability of the model for long-term sequence data. The signal is grouped by using the multi-head attention mechanism, which enables the model to focus on different features of the signal from multiple angles, improving the comprehensiveness and accuracy of feature extraction. Then, the grouping result is input into the LSTM network, and a residual connection is introduced, effectively solving the problems of gradient disappearance and gradient explosion, so that the model can more stably learn and extract long-term degradation features of industrial equipment. This method of constructing a wide-area perception path can deeply mine the running state changes of industrial equipment over a long period of time, and timely discover slow decline of equipment performance and potential fault hidden dangers.

[0026] In some embodiments of the present application, when the original signal is subjected to empirical mode decomposition to obtain a plurality of intrinsic mode function components and a residual term, the following calculation formula is used: ; Wherein, S(t) represents the input signal, N represents the number of intrinsic mode function components, IMFi(t) represents the i-th intrinsic mode function, and R(t) represents the residual term.

[0027] Specifically, to screen effective frequency bands, we perform one-dimensional convolution on each IMF component (Intrinsic Mode Function component) to extract time-domain features. After uniformizing the dimensions through pooling operations, the feature vectors are spliced and input into a fully connected network, and finally the normalized weights of each IMF are output. The Softmax function is used for weight calculation to ensure that the sum is 1. In the signal reconstruction process, only the IMF components with weight coefficients exceeding the preset threshold are retained. Threshold selection is an important means of noise suppression: retaining IMF components with weight coefficients exceeding the threshold, and suppressing the rest as noise. This preprocessing procedure significantly improves the proportion of key degradation frequency bands, providing high-quality input for subsequent time series modeling.

[0028] In some embodiments of the present application, when the reconstructed intrinsic mode function components are position encoded by a multi-frequency position encoding mechanism, the following calculation formula is used: ; Where PEj(k) represents the encoding value of position k in the jth dimension, M represents the total number of reconstructed intrinsic mode function components, corresponding to M independent frequency components, t represents the physical change period corresponding to the intrinsic mode function component (ensuring that the encoding period is consistent with the physical change trend), j represents the dimension index of position encoding (j = 1, 2,..., d, where d is the encoding dimension), and n is a positive integer.

[0029] It can be understood that the calculation formula of the multi-frequency position encoding mechanism accurately assigns position information to the reconstructed intrinsic mode function components. By introducing the physical change period t, the actual physical characteristics of the industrial equipment signal can be closely combined, so that the model can better capture the dynamic changes of the signal at different time scales. Different physical change periods reflect different change laws during the operation of the equipment, and this combination allows the model to more accurately understand the intrinsic characteristics of the signal. The setting of the dimension index j of position encoding and the positive integer n increases the flexibility and diversity of the encoding. Different dimension indexes can encode the signal from multiple angles, allowing the model to analyze and understand the characteristics of the signal from multiple levels. The presence of the positive integer n allows the complexity and accuracy of the encoding to be adjusted according to actual needs to adapt to the requirements of different industrial equipment and different application scenarios. This calculation formula takes into account the total number M of reconstructed intrinsic mode function components, which helps the model to grasp the scale and range of the signal as a whole. When dealing with large-scale industrial equipment signals, all relevant intrinsic mode function components can be considered more comprehensively to avoid information omission. In this way, the model can more accurately extract the long-term degradation characteristics of the industrial equipment, improving the accuracy and reliability of fault identification.

[0030] In some embodiments of the present application, when grouping is performed using a multi-head attention mechanism, the following calculation formula is used: ; where Q denotes the query matrix, K denotes the key matrix, V denotes the value matrix, dk denotes the dimension of the key vector, and softmax denotes the activation function.

[0031] Specifically, the periodic characteristics of the IMF components are embedded into the query (Q), key (K), and value (V) matrices using position encoding, thereby deeply integrating location information and physical periodic characteristics. To adapt to the multi-head attention framework, a periodic grouping strategy is adopted to divide all attention heads into M groups, where the i-th group is specifically responsible for processing the feature interaction corresponding to the i-th IMF component.

[0032] It can be understood that the calculation formula of the multi-head attention mechanism provides strong feature extraction capability for the processing of industrial equipment signals. Through the interaction of the query matrix Q, the key matrix K, and the value matrix V, the model can focus on different parts of the signal in different representation subspaces in parallel, thereby more comprehensively and meticulously capturing key information in the signal. The setting of the dimension dk of the key vector allows the model to adjust the granularity of attention to signal features according to actual needs, focusing on both local subtle features and overall macro features. The use of the activation function softmax allows the attention weights to be normalized, enabling the model to highlight important features and suppress irrelevant noise information when processing signals. This helps the model more accurately identify fault features in the signal and improves the accuracy of fault identification. In the grouping process, the multi-head attention mechanism can consider multiple different perspectives simultaneously, avoiding the limitations that may exist in a single attention mechanism. Multiple attention heads can analyze the signal from different angles, thereby discovering more hidden features and patterns. This multi-perspective analysis approach enables the model to more deeply understand the internal structure of the industrial equipment signal and improves the ability to identify fault patterns.

[0033] S3, a local domain-aware path is constructed, an LSTM-attention structure is adopted, high-order differentiation and sparse mask mechanisms are combined to process the original signal, and instantaneous fault features of the industrial equipment are extracted.

[0034] Specifically, accurately identifying local fault events is crucial for key areas such as equipment operation and maintenance and system safety, and signal analysis is a core means of capturing such events. Since local faults are usually accompanied by subtle mutations in physical states, these changes will exhibit specific feature patterns in the monitoring signal, and effectively extracting these feature patterns is a prerequisite for accurate detection.

[0035] In some embodiments of the present application, a local domain perception path is constructed, an LSTM-attention structure is adopted, and high-order differentiation and sparse mask mechanism are combined to process the original signal and extract instantaneous fault features of industrial equipment, including: normalizing the original signal and extracting differential features through high-order differentiation technology, the differential features including first-order differential features, second-order differential features and third-order differential features; inputting the differential features into an improved LSTM network, fusing original embedding features through a gating mechanism, constructing an attention mechanism for the differential features output by the LSTM network, and setting a sparse mask corresponding to the differential order to extract instantaneous fault features of industrial equipment.

[0036] It can be understood that the combination of the LSTM-attention structure, high-order differentiation and sparse mask mechanism provides an efficient and accurate method for extracting instantaneous fault features of industrial equipment. Normalizing the original signal can unify the signal to the same scale range, eliminate the dimensional differences between different signals, and make the subsequent feature extraction more stable and accurate. The use of high-order differentiation technology can deeply mine the rate information of the signal. First-order, second-order and third-order differential features reflect the dynamic characteristics of the signal from different dimensions, which helps to capture subtle changes at the moment of equipment failure. The improved LSTM network fuses original embedding features through a gating mechanism, which can effectively process sequence data and remember long-term dependencies in the signal. This feature enables the network to better track the trend of signal changes when processing dynamic signals of industrial equipment, and more comprehensively analyze fault features. Building an attention mechanism for the differential features output by the LSTM network can allow the model to automatically focus on important parts of the signal related to faults, enhancing the ability to capture key features. Setting a sparse mask corresponding to the differential order further improves the relevance of feature extraction, suppressing irrelevant information interference and highlighting features closely related to faults.

[0037] In some embodiments of the present application, when extracting differential features through high-order differentiation technology, the following calculation formula is used: ; Where △1S(t) represents the first-order differential feature, △2S(t) represents the second-order differential feature, and △3S(t) represents the third-order differential feature. S(t) represents the normalized original signal at time t, S(t+1) represents the normalized original signal at time t+1, S(t-1) represents the normalized original signal at time t-1, and S(t+2) represents the normalized original signal at time t+0.

[0038] Specifically, the first-order differential identifies progressive anomalies by detecting changes in signal slope; the second-order differential locates impact events by capturing acceleration changes; and the third-order differential extracts abrupt acceleration features and suppresses false impact interference. The differential operation is implemented using a causal convolution kernel to ensure real-time response of feature extraction.

[0039] It can be understood that the differential features are extracted by explicit calculation formula, so that the entire feature extraction process is more standardized and quantifiable. This not only facilitates the unified processing of different industrial equipment raw signals in practical applications, but also improves the repeatability and accuracy of feature extraction. The standardized calculation method reduces the interference of human factors and ensures the consistency of results in different data processing processes.

[0040] In some embodiments of the present application, when the above enhanced multi-dimensional feature vector is input into the LSTM module, the direct concatenation of LSTM and Transformer in the traditional cascading structure is easy to cause feature information attenuation. Therefore, the gating mechanism is improved to realize the dynamic fusion of LSTM hidden state and original embedding features. When the differential features are input into the improved LSTM network and the embedding features are fused through the gating mechanism, the following calculation formula is used; ; Where o represents the output result of the LSTM, f represents the embedding feature, cos(o, f) represents the cosine similarity between the LSTM output feature o and the embedding feature f, g represents the gating coefficient, sigmoid represents the activation function, and G(g1,..., gt) represents the output feature.

[0041] Specifically, the determination of the gating coefficient is based on the cosine similarity between the LSTM output feature and the embedding feature: when the similarity is too low, the weight of the embedding feature is increased to prevent information loss caused by LSTM feature degradation; when the similarity is higher, the contribution of the LSTM feature is enhanced. The gating function uses a Sigmoid activation function to ensure that the output value range is between [0, 1]. Through the improved gating mechanism, LSTM can effectively extract long-term and short-term time dependencies while preserving the key details of local fault events in multi-dimensional features (such as third-order differential features and second-order differential features). However, the capture of time correlation by LSTM focuses more on the dynamic dependence of the local neighborhood, which makes it difficult to model global correlations spanning long time periods (e.g., the association between faults at different time points, the cumulative effect of weak fault features over the entire time range, etc.). At this time, the attention mechanism module is introduced to take advantage of its ability to model long-range time correlations and supplement global information for the output features of the LSTM. By calculating the dependence weight between features at different time points, this module can explore the evolution law of local fault events in the complete time dimension.

[0042] It can be understood that the improved LSTM network and the fusion of embedded features through the gating mechanism greatly enhance the model's processing ability for sequence data. LSTM network itself is good at processing data with time series characteristics, and the improved LSTM network combined with a specific gating mechanism can more flexibly control the flow and storage of information. This process is realized through explicit calculation, making the entire fusion process more accurate and stable. The gating mechanism plays a key role in it, which can dynamically decide which information needs to be retained and which information needs to be forgotten according to the input embedded features. This allows the model to adaptively adjust the attention to different time periods and different features, better capturing long-term dependencies in the data. In the device fault identification scenario, the running state of the device often has certain continuity and correlation, and this effective capture of long-term dependencies can help the model more accurately identify fault patterns.

[0043] In some embodiments of the present application, a sparse mask corresponding to the differential order is set, including: setting a corresponding time window bandwidth for the differential feature; specifically: setting a first-order differential bandwidth 3z for the first-order differential feature, a second-order differential bandwidth 2z for the second-order differential feature, and a third-order differential bandwidth 1z for the third-order differential feature; wherein z is a predetermined constant.

[0044] Specifically, the attention mechanism of the standard Transformer will introduce irrelevant time feature interference in the local path. To solve this problem, attention mechanisms are constructed for the first-order, second-order, and third-order differential features processed by LSTM, and corresponding differential masks are set. The specific process is as follows: based on the time scale characteristics of different order differential features (gradual features require long-term correlation, and instantaneous effects require focusing on local areas), attention sparsification is realized through time window constraints: set a dedicated bandwidth W for different order features, and only calculate the correlation weight of the time steps within the window, and shield irrelevant information outside the window. This design not only simplifies the calculation process, but also improves the feature specificity. The bandwidth of the third-order differential is fixed as a constant value z, and the bandwidth parameters of the first-order and second-order differentials are designed accordingly. The first-order differential (W=3z) focuses on gradual abnormalities (such as device performance degradation), which have long-term persistence. This method needs to correlate the time features before and after 3z steps to capture the cumulative effect of the trend, and a wide time window ensures that slow-changing overall patterns are not missed. The second-order differential (W=2z) is mainly composed of transient signals and needs to focus on the core area of 2z steps around the peak. A narrow time window can avoid irrelevant time interference and enhance the recognition of impact features. The third-order differential (W=z) relies on instantaneous mutation characteristics and only needs to focus on the features of the current z time step. This method can filter cross-time noise to the maximum extent and highlight the instantaneous differences of real impacts. Finally, the third-order attention is calculated by combining with the differential mask, and the feature splicing is performed before the output layer.

[0045] It can be understood that by setting corresponding sparse masks and time window bandwidths for different differential orders of features, device operation data can be processed more finely. Different order differential features reflect changes in the device operation state to different degrees, and setting appropriate bandwidths for them can enable the model to more specifically capture the key information carried by each order feature.

[0046] S4, dimensionally aligning the long-term degradation features of the wide-area perception path with the instantaneous fault features of the local-area perception path, and outputting a fault recognition result of the industrial equipment.

[0047] In some embodiments of the present application, dimensionally aligning the long-term degradation features of the wide-area perception path with the instantaneous fault features of the local-area perception path includes: setting a hybrid loss function, the hybrid loss function including a wide-area loss function L1 and a local-area loss function L2, an expression of the hybrid loss function being: L = aL1 + (1-a)L2, where L represents a total loss function, and a represents a dynamic adjustment coefficient; and achieving collaborative optimization of the wide-area perception path and the local-area perception path through the hybrid loss function.

[0048] Specifically, the core of the hybrid loss function is to design a sub-loss function for each of the two parallel networks. For the wide-area perception path, its sub-loss function L1 is used to constrain the learning accuracy of long-term trends; for the local-area perception path, its sub-loss function L2 needs to focus on learning local features. a is a dynamic adjustment coefficient. In the early stage of training, a can be set to 0.5 to balance the learning effects of the two paths; as the training progresses, a will be dynamically adjusted according to the descending rates of the two sub-loss functions on the validation set. If L1 has a faster descending rate, it means that the learning effect of the wide-area perception path is better, and at this time the value of a will be increased to enhance its learning weight; otherwise, the value of a will be decreased to guide the local-area perception path to play a greater role. This dynamic adjustment mechanism avoids the learning bias that may be caused by fixed weights, and ensures that the two networks always maintain a complementary relationship during the training process. Through the above hybrid loss function, the wide-area perception path and the local-area perception path can both be optimized in their respective professional feature dimensions, and collaborative learning is achieved through dynamic weights. This mechanism can effectively improve the accuracy of the recognition result in capturing long-term trends and local details.

[0049] It can be understood that the synergistic optimization of the wide-area perception path and the local-area perception path is achieved by setting the hybrid loss function, which can fully utilize the advantages of both in equipment fault identification. The long-term degradation features of the wide-area perception path can reflect the performance changes and potential fault risks of the equipment over a long period of time, while the instantaneous fault features of the local-area perception path can capture sudden abnormal situations during the operation of the equipment. By dimensionally aligning and synergistically optimizing these two features, the model can take into account both the long-term running state and the short-term sudden failure of the equipment.

[0050] In equipment fault identification, many faults do not occur suddenly, but have a process from quantitative change to qualitative change. The long-term degradation features of the wide-area perception path can help us discover the gradual decline trend of the performance of the equipment in advance, such as wear and tear, aging, and other problems of the equipment parts. The instantaneous fault features of the local-area perception path can timely alarm when the equipment has a sudden failure, such as short circuit, open circuit, and other situations. By combining the two through the hybrid loss function, the model can more sensitively capture the instantaneous fault features on the basis of discovering the long-term degradation features, thereby improving the accuracy and timeliness of fault identification.

[0051] The introduction of the dynamic adjustment coefficient a enables the model to flexibly adjust the weights of the wide-area loss function L1 and the local-area loss function L2 according to the actual situation. In different application scenarios, the running characteristics and fault modes of the equipment may be different. For example, in some scenarios where the stability of the equipment is required to be high, more attention may be paid to the long-term degradation features of the wide-area perception path, and in this case, the value of a can be appropriately increased; while in some scenarios where the response speed of the fault is required to be high, more attention may be paid to the instantaneous fault features of the local-area perception path, and in this case, the value of a can be appropriately reduced. This dynamic adjustment method enhances the adaptability and flexibility of the model, making it better cope with various complex and variable actual situations.

[0052] As shown in Figure 2 On the other hand, the application also discloses an equipment fault identification system based on a dual-domain perception path, which is used to apply the equipment fault identification method based on the dual-domain perception path, and comprises: A signal acquisition module configured to monitor an industrial equipment in real time and acquire original signals of the industrial equipment.

[0053] A wide-area perception path construction module configured to construct a wide-area perception path, adopt an attention-LSTM structure, process the original signals by combining an empirical mode decomposition and an improved multi-frequency position encoding mechanism, and extract long-term degradation features of the industrial equipment.

[0054] The local domain perception path construction module is configured to construct a local domain perception path, adopts an LSTM-attention structure, processes the original signal by combining a high-order differential and a sparse mask mechanism, and extracts instantaneous fault features of the industrial equipment.

[0055] The fault identification output module is configured to perform dimension alignment on long-term degradation features of the wide-area perception path and instantaneous fault features of the local domain perception path, and output a fault identification result of the industrial equipment.

[0056] Through the cooperative work of the various modules, the present application realizes efficient and accurate identification of the faults of the industrial equipment. The signal acquisition module ensures that the original signal of the equipment can be acquired in real time, providing a data basis for subsequent analysis. The wide-area perception path construction module extracts long-term degradation features by using advanced structures and mechanisms, which helps to discover potential fault hidden dangers of the equipment in advance. The local domain perception path construction module can timely capture instantaneous fault features of the equipment, providing support for rapid response to faults. The fault identification output module performs dimension alignment on the two kinds of features and outputs the result, so that the system can comprehensively consider the long-term running state and short-term burst situation of the equipment, greatly improving the accuracy and reliability of fault identification. Moreover, the modular design makes the system have good scalability and maintainability. If the system needs to be upgraded or optimized, the specific module can be improved, without affecting the normal operation of other modules.

[0057] Those skilled in the art will understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can adopt a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0058] The present application is described with reference to flowcharts and / or block diagrams according to the methods, devices (systems) and computer program products of the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The device that implements the functions specified in one flow or multiple flows and / or blocks Figure 1 The device that implements the functions specified in one flow or multiple flows and / or blocks

[0059] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the flow Figure 1 one or more flows and / or blocks Figure 1 one or more blocks or multiple blocks.

[0060] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the flow Figure 1 one or more flows and / or blocks Figure 1 one or more blocks or multiple blocks.

[0061] Finally, it should be noted that the above-mentioned embodiments are merely used to illustrate the technical solutions of the present application, rather than limiting the same. Even though the present application has been described in detail with reference to the above-mentioned embodiments, those skilled in the art should understand that the specific embodiments of the present application can be modified or replaced equivalently, and any modification or replacement without departing from the spirit and scope of the present application should be covered within the protection scope of the claims of the present application.

Claims

1. A device fault identification method based on dual-domain sensing path, characterized in that, include: Real-time monitoring of industrial equipment to obtain raw signals from the equipment; A wide-area sensing path is constructed, and an attention-LSTM structure is adopted. The original signal is processed by combining empirical mode decomposition and an improved multi-frequency position coding mechanism to extract the long-term degradation characteristics of industrial equipment. A local domain sensing path is constructed, and an LSTM-attention structure is used to process the original signal by combining high-order differentiation and sparse masking mechanisms to extract the instantaneous fault features of industrial equipment. By dimensionally aligning the long-term degradation characteristics of the wide-area sensing path with the instantaneous fault characteristics of the local-area sensing path, the fault identification results of industrial equipment are output.

2. The equipment fault identification method based on dual-domain sensing path according to claim 1, characterized in that, A wide-area sensing path is constructed, employing an attention-LSTM structure, combined with empirical mode decomposition and an improved multi-frequency position coding mechanism to process the original signal, extracting long-term degradation features of industrial equipment, including: The original signal is subjected to empirical mode decomposition to obtain several intrinsic mode function components and residual terms; For each intrinsic mode function component, feature extraction and weight calculation are performed, and intrinsic mode function components with weight coefficients exceeding a preset threshold are selected for signal reconstruction; The reconstructed intrinsic mode function components are position-coded using a multi-frequency position coding mechanism. A multi-head attention mechanism is used for grouping, and the grouping results are input into an LSTM network. Residual connections are introduced to extract long-term degradation characteristics of industrial equipment.

3. The equipment fault identification method based on dual-domain sensing path according to claim 2, characterized in that, When performing position encoding on the reconstructed intrinsic mode function components using a multi-frequency position encoding mechanism, the following calculation formula is used; ； Where PEj(k) represents the encoded value at position k in the j-th dimension, M represents the total number of reconstructed intrinsic mode function components, t represents the physical change period corresponding to the intrinsic mode function component, j represents the dimension index of the position encoding, and n is a positive integer.

4. The equipment fault identification method based on dual-domain sensing path according to claim 2, characterized in that, When using a multi-head attention mechanism for grouping, the following calculation formula is used; ； Where Q represents the query matrix, K represents the key matrix, V represents the value matrix, dk represents the dimension of the key vector, and softmax represents the activation function.

5. The equipment fault identification method based on dual-domain sensing path according to claim 1, characterized in that, A local domain sensing path is constructed, employing an LSTM-attention structure, combined with high-order differentiation and sparse masking mechanisms to process the original signal, extracting instantaneous fault features of industrial equipment, including: The original signal is normalized, and differential features are extracted using higher-order differential techniques. These differential features include first-order differential features, second-order differential features, and third-order differential features. The differential features are input into an improved LSTM network, and the original embedded features are fused through a gating mechanism. An attention mechanism is constructed on the differential features output by the LSTM network, and a sparse mask corresponding to the differential order is set to extract the instantaneous fault features of industrial equipment.

6. The equipment fault identification method based on dual-domain sensing path according to claim 5, characterized in that, When extracting differential features using higher-order differential techniques, the following calculation formula is used; ； Where △1S(t) represents the first-order differential feature, △2S(t) represents the second-order differential feature, △3S(t) represents the third-order differential feature, S(t) represents the original signal after normalization at time t, S(t+1) represents the original signal after normalization at time t+1, S(t-1) represents the original signal after normalization at time t-1, and S(t+2) represents the original signal after normalization at time t+0.

7. The equipment fault identification method based on dual-domain sensing path according to claim 5, characterized in that, When the differential features are input into the improved LSTM network and the embedded features are fused through a gating mechanism, the following calculation formula is used; ； Where o represents the output of LSTM, f represents the embedded feature, cos(o,f) represents the cosine similarity between the LSTM output feature o and the embedded feature f, g represents the gating coefficient, sigmoid represents the activation function, and G(g1,...,gt) represents the output feature.

8. The equipment fault identification method based on dual-domain sensing path according to claim 5, characterized in that, Setting a sparse mask corresponding to the differential order includes: setting the corresponding time window bandwidth for the differential feature; specifically: Set a first-order differential bandwidth of 3z for the first-order differential feature, a second-order differential bandwidth of 2z for the second-order differential feature, and a third-order differential bandwidth of 1z for the third-order differential feature; where z is a preset constant.

9. The equipment fault identification method based on dual-domain sensing path according to claim 1, characterized in that, Aligning the long-term degradation characteristics of wide-area sensing paths with the instantaneous fault characteristics of local-area sensing paths in terms of dimensions, including: A hybrid loss function is defined, which includes a wide-area loss function L1 and a local-area loss function L2. The expression of the hybrid loss function is: L = αL1 + (1-α)L2, where L represents the total loss function and α represents the dynamic adjustment coefficient. The hybrid loss function enables the collaborative optimization of wide-area perception paths and local-area perception paths.

10. A device fault identification system based on dual-domain sensing path, used to apply the device fault identification method based on dual-domain sensing path as described in any one of claims 1-9, characterized in that, include: The signal acquisition module is configured to monitor industrial equipment in real time and acquire the raw signals of the industrial equipment; The wide-area sensing path construction module is configured to construct a wide-area sensing path. It adopts an attention-LSTM structure and combines empirical mode decomposition and an improved multi-frequency position coding mechanism to process the original signal and extract the long-term degradation characteristics of industrial equipment. The local domain sensing path construction module is configured to construct a local domain sensing path. It uses an LSTM-attention structure and combines high-order differentiation and sparse masking mechanisms to process the original signal and extract the instantaneous fault features of industrial equipment. The fault identification output module is configured to dimensionally align the long-term degradation characteristics of the wide-area sensing path with the instantaneous fault characteristics of the local-area sensing path, and output the fault identification results of the industrial equipment.

Images