Industrial equipment fault prediction and diagnosis system based on deep learning

Abstract

The present invention relates to the technical field of industrial equipment fault monitoring and diagnosis, and specifically to a deep learning-based industrial equipment fault prediction and diagnosis system, comprising a dynamic data alignment module, a multidimensional feature fusion module, a fault mode decoupling module, and an interpretable diagnostic mapping module. The system comprises: a dynamic data alignment module for constructing a multidimensional equipment status data stream under a unified timeline; a multidimensional feature fusion module for performing orthogonal decomposition in feature space to extract a set of independent feature vectors; a fault mode decoupling module for constructing a fault mode space projection template; and an interpretable diagnostic mapping module for performing constraint matching to generate a visual diagnostic report. The present invention achieves accurate fault location and remaining life prediction by dynamically aligning and multidimensional feature fusion of multi-source heterogeneous sensor data, combining fault mode decoupling and interpretable diagnostic mapping, while reducing feature redundancy and noise interference.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial equipment fault monitoring and diagnosis, and in particular to an industrial equipment fault prediction and diagnosis system based on deep learning. Background Art

[0002] With the continuous improvement of the intelligence level of industrial equipment, the deployment of sensor networks in equipment operation status monitoring is becoming increasingly intensive. By collecting multi-source physical quantities such as vibration, temperature, current, voltage, and acoustic emission, real-time perception and remote management of equipment status are achieved; relying on these multi-source status parameters, artificial intelligence-based fault identification technology is gradually applied to fault prediction and health management scenarios of complex equipment, especially in industries such as wind power, petrochemicals, electric power, and heavy machinery, showing certain application potential.

[0003] However, existing systems mostly rely on manually constructed features or single-source data, and are unable to effectively integrate information from time, space, and multi-dimensional features. They also lack deep decoupling and mapping mechanisms at the physical structure level, resulting in problems such as unclear fault location, poor interpretability, and fragmented diagnostic reports. These problems seriously restrict the promotion and application of the system in engineering. Therefore, there is an urgent need for an industrial equipment fault prediction and diagnosis system based on deep learning to solve the above problems. Summary of the Invention

[0004] Based on the above objectives, the present invention provides an industrial equipment fault prediction and diagnosis system based on deep learning.

[0005] The deep learning-based industrial equipment fault prediction and diagnosis system includes a dynamic data alignment module, a multi-dimensional feature fusion module, a fault mode decoupling module, and an explainable diagnostic mapping module.

[0006] Dynamic data alignment module: used to align the multi-source heterogeneous sensor data of industrial equipment with temporal and spatial constraints, and build a multi-dimensional equipment status data stream under a unified timeline;

[0007] Multi-dimensional feature fusion module: used to receive the aligned multi-dimensional device status data stream, perform orthogonal decomposition in the feature space, and extract a set of independent feature vectors;

[0008] Fault mode decoupling module: constructs a fault mode space projection template based on a set of independent feature vectors, and outputs a decoupled feature matrix containing fault type codes and fault severity levels;

[0009] Explainable diagnostic mapping module: used to constrain matching of the decoupled feature matrix with the equipment physical parameter library to generate a visual diagnostic report containing a fault location map and remaining life prediction.

[0010] Optionally, the dynamic data alignment module includes a data acquisition unit, a time synchronization unit, a spatial registration unit, and a data interpolation unit; wherein:

[0011] Data acquisition unit: used to synchronously collect original state parameter data from vibration sensors, temperature sensors, current and voltage sensors, and bearing acoustic emission sensors deployed on industrial equipment, including vibration acceleration values, surface temperature values, current values, voltage values, and acoustic emission signal amplitudes, and to attach original timestamps and 3D installation coordinates to each type of sensor data;

[0012] Time synchronization unit: used to receive the state parameter data output by the data acquisition unit, align the timestamps of all state parameters using a sliding time window based on the set master time reference, and output a state data sequence with consistent time index;

[0013] Spatial registration unit: It is used to receive the time-synchronized state data sequence, construct a spatial mapping function based on the three-dimensional installation coordinates of each sensor, and uniformly convert the position vectors of the state parameters to form a spatially consistent multi-dimensional device state data stream.

[0014] Optionally, the spatial registration unit includes:

[0015] Receive the three-dimensional installation coordinate information corresponding to each type of status parameter ,in Indicates the sensor number, Respectively represent The three-dimensional spatial position of each sensor in the industrial equipment coordinate system;

[0016] Construct a local spatial reference frame R based on the structural geometry model of the industrial equipment, define the global alignment direction based on R, and transform all sensor coordinates to relative coordinates under the reference frame R. ;

[0017] The aligned sensor coordinates Combined with the state parameter value at the corresponding moment, a unified format of spatial state pair is constructed. ;

[0018] The state vectors of all sensors at the same time index By number Aggregate and output a spatially consistent multi-dimensional device status data stream, expressed as: ,in, is the total number of sensors, For time State frame after lower spatial registration.

[0019] Optionally, the multidimensional feature fusion module includes a state matrix construction unit, a feature correlation determination unit, and an orthogonal feature extraction unit; wherein:

[0020] State matrix construction unit: used to receive the unified time axis device state data stream output by the dynamic data alignment module, and align the spatial state of each sensor in each time slice. Converted into row vector form, stacked in time series to form the original state matrix, each column of the state matrix corresponds to a state parameter, and each row is a snapshot of the spatial state at a single time point;

[0021] Feature correlation discrimination unit: used to perform correlation analysis on the state parameters of each column in the original state matrix. Through joint statistical indicators based on time and space dimensions, it identifies parameter pairs with multicollinearity or redundant coupling and marks their potential coupling structure.

[0022] Orthogonal feature extraction unit: It is used to perform orthogonal decomposition processing based on feature correlation judgment, reconstruct a set of independent feature vectors without correlation interference, and retain the physical meaning mapping relationship between the original parameters and the extracted features during the decoupling process to achieve structural restoration with physical interpretability.

[0023] Optionally, the orthogonal feature extraction unit includes:

[0024] Covariance matrix construction subunit: Construct the covariance matrix based on the state parameter set retained after feature correlation discrimination , used to characterize the linear relationship structure between various parameters;

[0025] Feature orthogonalization subunit: used for covariance matrix Perform eigenvalue decomposition, extract the corresponding eigenvector set, and map the original state parameter set to the orthogonal feature space through projection transformation;

[0026] Physical mapping preservation subunit: used to record the projection structure of each orthogonal eigenvector in the original state parameter space, and form an orthogonal feature-original parameter mapping comparison table by calculating the feature contribution rate and reverse mapping the dominant physical parameters corresponding to each independent feature.

[0027] Optionally, the fault mode decoupling module includes a fault template construction unit, a mode projection execution unit, and a fault level determination unit; wherein:

[0028] Fault template construction unit: Based on the identified independent feature vector set and the characteristic response patterns of different types of faults in the historical fault sample set, typical fault feature clusters are extracted and organized into standardized fault pattern templates. The corresponding relationship between each fault type and its typical feature expression is established;

[0029] Pattern projection execution unit: used to align and match the independent feature vector set extracted at the current moment or in the current cycle with the constructed fault pattern template. By calculating the similarity or distance index in the feature space, it determines which type of fault template the current feature distribution is closer to, and generates a fault type code with a unique identifier.

[0030] Fault level discrimination unit: Based on the dimensions of amplitude change, temporal stability, and spatial aggregation of the feature vector, combined with the risk level standard corresponding to the fault type, the severity of the identified fault instance is evaluated, and the classification result is output. Structured data of fault type code + fault severity level is generated to form the final decoupled feature matrix.

[0031] Optionally, the fault template construction unit includes:

[0032] Feature label grouping subunit: used to group the data with labeled fault types in the historical fault sample set. Each group corresponds to a known fault type, and extracts the independent feature vector set within the corresponding time period to construct multiple sets of feature vector sequences indexed by fault type.

[0033] Feature clustering analysis subunit: used to perform feature morphology clustering analysis on feature vectors using an unsupervised clustering algorithm under each fault type group, classifying feature vectors with similar spatial distribution into the same feature cluster, and extracting the most representative k feature clusters of each fault type as its typical feature cluster;

[0034] Template standardization generation subunit: used to construct the average vector of the typical feature cluster of each fault type as the standardized template expression of this type of fault, and establish a one-to-one mapping structure between fault type number and feature template vector to form a fault template database.

[0035] Optionally, the fault level determination unit includes:

[0036] Amplitude change calculation subunit: used to perform difference analysis on the numerical amplitude of each feature dimension in the currently identified independent feature vector set and the corresponding fault template vector, obtain the feature offset, and calculate the fault offset coefficient based on the total amount of full-dimensional amplitude change , used to reflect the intensity characteristics of fault manifestation;

[0037] Time stability evaluation subunit: used to analyze the change trend of independent feature vectors in N consecutive time slices, and calculate the change fluctuation value using the mean square error indicator , used to express the time-continuous stability of fault characteristics;

[0038] Spatial aggregation analysis subunit: used to determine the distribution concentration of the current fault characteristics in the sensor space coordinates, based on The distance between the centroid and the location of the fault impact point , used to represent the feature space focality;

[0039] Level determination output subunit: used to , , The three indicators are normalized and weighted to form a severity assessment score. The fault severity level is output according to the scoring interval in the risk level standard table corresponding to the fault type. Finally, it is combined with the determined fault type code into structured data and embedded in the decoupled feature matrix under the current time index.

[0040] Optionally, the explainable diagnosis mapping module includes a feature-structure mapping unit, a fault location generation unit, and a life prediction output unit; wherein:

[0041] Feature-Structure Mapping Unit: This unit receives a decoupled feature matrix containing fault type codes and severity levels, extracts basic attributes such as structural component numbers, material categories, key component coordinates, and stress boundary conditions from the equipment physical parameter library, and establishes a one-to-one constraint mapping between feature dimensions and the equipment's physical structure based on the historical correspondence between fault types and dominant state parameters, identifying structural regions associated with current feature behaviors.

[0042] Fault location generation unit: This unit is used to extract the spatial location of the physical part corresponding to the current fault based on the mapping results. It then generates a structural map highlighting the current diagnosis results based on the sensor's 3D installation coordinates and the structural partition map. It then outputs the fault location map and embeds it into a visual diagnostic report in the form of structural number, component name, and coordinate position.

[0043] Life prediction output unit: Based on the evolution trajectory of the same type of faults in historical degradation data, combined with the current fault severity level and its time stability index, the remaining life statistical model of the corresponding component in the physical parameter library is retrieved, and the remaining life prediction value RUL is output. It is presented as a quantitative indicator together with the fault location map in the final visual diagnosis report.

[0044] Optionally, the feature-structure mapping unit includes:

[0045] Parameter association analysis subunit: This subunit is used to extract the dominant state parameter set for each fault type from the established fault template database, query the device physical parameter library for structural parameter records that have a one-to-one response relationship with these state parameters, and construct an initial mapping set between feature dimensions and structural attributes.

[0046] Structural constraint matching subunit: used to introduce the amplitude weights of each feature dimension in the current decoupled feature matrix based on the initial mapping set and establish a constraint weight vector;

[0047] Regional correlation identification subunit: used to project the spatial coordinate information in the structural attribute set to the three-dimensional device model, aggregate high-response structural points based on the distribution density of the constraint weight vector on the spatial coordinates, and calculate the spatial response concentration ,when Exceeding the set threshold , then the region is determined to be the target related structure region.

[0048] Beneficial effects of the present invention:

[0049] The present invention, by introducing two core links, dynamic data alignment and multi-dimensional feature fusion, achieves the precise unification of multi-source heterogeneous sensor data in time and space dimensions, overcoming the fault identification bias problem caused by data asynchrony and parameter redundancy in existing solutions; the organic connection between fault mode decoupling and explainable diagnostic mapping makes independent feature vectors correspond to specific physical structures, greatly improving the accuracy of fault location and avoiding the disadvantage of difficulty in physical tracing under pure data drive.

[0050] The present invention realizes a coherent analysis process from feature extraction, fault identification to life prediction through fault template management and real-time severity classification based on historical degradation laws, providing visual diagnostic reports and remaining life indicators for industrial sites. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0052] Figure 1 A schematic diagram of an industrial equipment fault prediction and diagnosis system according to an embodiment of the present invention;

[0053] Figure 2 Schematic diagram of a fault mode decoupling module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0054] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. It is also noted that, to provide a more detailed description, the following embodiments are best and preferred embodiments, and those skilled in the art may employ alternative methods for implementing certain known technologies. Furthermore, the accompanying drawings are intended only to provide a more detailed description of the embodiments and are not intended to limit the present invention.

[0055] It should be noted that references in the specification to "one embodiment," "an embodiment," "exemplary embodiments," "some embodiments," etc. indicate that the described embodiments may include specific features, structures, or characteristics, but not necessarily every embodiment will include such specific features, structures, or characteristics. Furthermore, when specific features, structures, or characteristics are described in conjunction with an embodiment, it is within the knowledge of persons skilled in the relevant art to implement such features, structures, or characteristics in conjunction with other embodiments (whether or not explicitly described).

[0056] In general, terms can be understood, at least in part, from their use in context. For example, depending at least in part on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular sense, or can be used to describe a combination of features, structures, or characteristics in the plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey an exclusive set of factors, but can instead, depending at least in part on the context, allow for the presence of other factors that are not necessarily explicitly described.

[0057] like Figure 1-Figure 2 As shown in FIG, the deep learning-based industrial equipment fault prediction and diagnosis system includes a dynamic data alignment module, a multi-dimensional feature fusion module, a fault mode decoupling module, and an explainable diagnosis mapping module; wherein:

[0058] Dynamic data alignment module: used to align the multi-source heterogeneous sensor data of industrial equipment with temporal and spatial constraints, and build a multi-dimensional equipment status data stream under a unified timeline;

[0059] Multi-dimensional feature fusion module: used to receive the aligned multi-dimensional device status data stream, perform orthogonal decomposition in the feature space, and extract a set of independent feature vectors;

[0060] Fault mode decoupling module: constructs a fault mode space projection template based on a set of independent feature vectors, and outputs a decoupled feature matrix containing fault type codes and fault severity levels;

[0061] Explainable diagnostic mapping module: used to constrain matching of the decoupled feature matrix with the equipment physical parameter library to generate a visual diagnostic report containing a fault location map and remaining life prediction.

[0062] The dynamic data alignment module includes a data acquisition unit, a time synchronization unit, a spatial registration unit, and a data interpolation unit; wherein:

[0063] Data acquisition unit: used to synchronously collect original state parameter data from vibration sensors, temperature sensors, current and voltage sensors, and bearing acoustic emission sensors deployed on industrial equipment, including vibration acceleration values, surface temperature values, current values, voltage values, and acoustic emission signal amplitudes, and to attach original timestamps and 3D installation coordinates to each type of sensor data;

[0064] Time synchronization unit: used to receive the state parameter data output by the data acquisition unit, align the timestamps of all state parameters using a sliding time window based on the set master time reference, and output a state data sequence with consistent time index;

[0065] Spatial registration unit: This unit receives the time-synchronized state data sequence, constructs a spatial mapping function based on the three-dimensional installation coordinates of each sensor, and uniformly converts the position vectors of the state parameters to form a spatially consistent multi-dimensional device state data stream. Through the collaborative processing of the above units, the dynamic data alignment module implements unified time and space constraints on multi-source state parameters such as vibration, temperature, electrical energy, and acoustic emission, constructs a multi-dimensional device state data stream under a unified time axis, and provides a consistent and standardized input data structure for the subsequent feature fusion process.

[0066] The spatial registration unit includes:

[0067] Receive the three-dimensional installation coordinate information corresponding to each type of status parameter ,in Indicates the sensor number, Respectively represent The three-dimensional spatial position of each sensor in the industrial equipment coordinate system;

[0068] Construct a local spatial reference frame R based on the structural geometry model of the industrial equipment, define the global alignment direction based on R, and transform all sensor coordinates into relative coordinates under the reference frame R. , the following linear transformation formula is used for coordinate mapping: ,in, is the original coordinate column vector, is the coordinate column vector after mapping, for The rotation matrix of for The translation vector of

[0069] The aligned sensor coordinates Combined with the state parameter value at the corresponding moment, a unified format of spatial state pair is constructed. , the expression is: ,in, Indicates time Time sensor Corresponding state parameters; Represents the single-point multidimensional state vector after completing the spatial mapping;

[0070] The state vectors of all sensors at the same time index By number Aggregate and output a spatially consistent multi-dimensional device status data stream, expressed as: ,in, is the total number of sensors, For time The state frame after spatial registration has unified coordinate representation and structural alignment characteristics; the above steps construct a spatial reference frame based on the geometric structure of industrial equipment, combine the linear mapping model to uniformly transform the sensor position, and bind the converted position to the state parameter to ensure the consistency of multi-source sensor data in the spatial dimension, provide a stable and structurally unified spatial foundation for the subsequent feature extraction process, and effectively improve the spatial accuracy and physical correspondence of fault feature identification.

[0071] The multi-dimensional feature fusion module includes a state matrix construction unit, a feature correlation judgment unit and an orthogonal feature extraction unit; wherein:

[0072] State matrix construction unit: used to receive the unified time axis device state data stream output by the dynamic data alignment module, and align the spatial state of each sensor in each time slice. Converted into row vector form, stacked in time series to form the original state matrix, each column of the state matrix corresponds to a state parameter, and each row is a snapshot of the spatial state at a single time point;

[0073] Feature correlation discrimination unit: used to perform correlation analysis on the state parameters of each column in the original state matrix. Through joint statistical indicators based on time and space dimensions, it identifies parameter pairs with multicollinearity or redundant coupling and marks their potential coupling structure.

[0074] The feature correlation discrimination unit includes:

[0075] Time correlation identification subunit: used to identify any two columns of state parameter vectors in the state matrix The Pearson correlation coefficient is calculated for the synergistic change trend on the entire time axis to identify parameter pairs with strong time correlation. The calculation formula is as follows:

[0076] ,in, is the length of the time series and They are and The mean of Greater than threshold (e.g. 0.85), it is determined to be a time-collinear parameter pair;

[0077] Spatial co-variation identification subunit: used for Under this condition, the coordinated change relationship of multiple state parameters in space is judged and the parameters are calculated. and Gather at sensor location The spatial structure similarity on , according to the spatial position index of each sensor , construct the following point pair covariance model:

[0078] ,in, is the total number of sensors, and Respectively represent The state value of a position, if Continuously greater than the space threshold , then it is marked as a spatial redundant pair;

[0079] Redundant coupling marking subunit: It is used to perform intersection screening on the parameter pairs obtained through temporal correlation identification and spatial co-variation identification, extract the state parameter pairs that show strong coupling relationships in both temporal and spatial dimensions, and mark them as redundant coupling structures as feature compression objects in subsequent orthogonal decomposition processing.

[0080] Orthogonal feature extraction unit: used to perform orthogonal decomposition processing based on feature correlation judgment, reconstruct a set of independent feature vectors without correlation interference, and retain the physical meaning mapping relationship between the original parameters and the extracted features during the decoupling process to achieve structural restoration of physical interpretability; the structure of the above unit realizes the extraction of low-redundancy and highly representative feature expressions from the original multi-dimensional state parameters by constructing a state matrix, judging the correlation between parameters, and extracting a set of independent feature vectors with a physical interpretation basis, effectively reducing parameter interference and noise propagation in the subsequent modeling process, and improving the accuracy and traceability of fault identification results.

[0081] The orthogonal feature extraction unit includes:

[0082] Covariance matrix construction subunit: Construct the covariance matrix based on the state parameter set retained after feature correlation discrimination , which is used to characterize the linear relationship structure between various parameters. Its calculation formula is: ,in, for The state parameter vector at time , is the mean vector of state parameters in the time dimension, is the length of the time series, The dimension is A symmetric positive definite matrix, is the number of state parameters involved in the analysis;

[0083] Feature orthogonalization subunit: used for covariance matrix Perform eigenvalue decomposition and extract the corresponding eigenvector set , and the original state parameter set is mapped to the orthogonal feature space through projection transformation. The formula is as follows: ,in, for An orthogonal matrix, each column represents an independent feature in one direction, It is a feature representation matrix in orthogonal space, with the properties that the columns are linearly independent and the covariance is zero;

[0084] Physical mapping preservation subunit: This unit records the projection structure of each orthogonal feature vector in the original state parameter space. By calculating the feature contribution rate and reversely mapping the dominant physical parameters corresponding to each independent feature, it forms an orthogonal feature-original parameter mapping comparison table, providing parameter support for physical interpretability in the subsequent fault decoupling process.

[0085] The specific calculation steps are as follows:

[0086] First, the contribution rate of each eigenvector in the orthogonalization process to each parameter in the original state parameter space is calculated, specifically the contribution rate of each orthogonal feature to the original parameter is calculated. , the calculation formula is: ,in, Represents the feature vector Middle elements (i.e., the original parameters The projection coefficient in the feature direction), The larger the value, the more significant the impact of the original parameter on the orthogonal feature;

[0087] Then, in each orthogonal feature, select one or more original parameters with the largest contribution rate and set the contribution rate threshold ,when: , then the original parameters are considered to be orthogonal features The dominant physical parameters of

[0088] Finally, each orthogonal feature identified above is matched with its dominant physical parameter to generate an orthogonal feature-original parameter mapping comparison table, which records the orthogonal feature number, the corresponding dominant original parameter and the contribution rate range.

[0089] Table 1 Orthogonal feature-original parameter mapping comparison:

[0090] ;

[0091] Through covariance modeling, feature space orthogonalization and physical mapping reconstruction, the above-mentioned subunits achieve effective dimensionality reduction and decoupling of the original redundant state parameters, construct independent feature expressions with clear physical meanings, effectively improve the accuracy and interpretability of subsequent fault mode identification, and reduce the impact of multicollinearity on model stability.

[0092] The fault mode decoupling module includes a fault template construction unit, a mode projection execution unit and a fault level judgment unit; wherein:

[0093] Fault template construction unit: Based on the identified independent feature vector set and the characteristic response patterns of different types of faults in the historical fault sample set, typical fault feature clusters are extracted and organized into standardized fault pattern templates. The corresponding relationship between each fault type and its typical feature expression is established;

[0094] Pattern projection execution unit: used to align and match the independent feature vector set extracted at the current moment or in the current cycle with the constructed fault pattern template. By calculating the similarity or distance index in the feature space, it determines which type of fault template the current feature distribution is closer to, and generates a fault type code with a unique identifier.

[0095] Fault level discrimination unit: Based on the dimensions of amplitude change, temporal stability, and spatial aggregation of the feature vector, combined with the risk level standard corresponding to the fault type, the severity of the identified fault instances is assessed and the classification results are output. Structured data of fault type code + fault severity level are generated to form the final decoupled feature matrix. The above-mentioned unit structure spatially maps independent features with standard fault templates to identify the fault mode closest to the current working condition, and combines multi-dimensional statistical features to determine the fault level, constructing a decoupled feature matrix with standardized structure and clear expression, providing an accurate and controllable decision-making basis for subsequent diagnostic output, while enhancing the model's portability and interpretability.

[0096] The fault template building block includes:

[0097] Feature label grouping subunit: used to group the data with labeled fault types in the historical fault sample set. Each group corresponds to a known fault type, and extracts the independent feature vector set within the corresponding time period to construct multiple sets of feature vector sequences indexed by fault type.

[0098] Feature clustering analysis subunit: used to perform feature morphology clustering analysis on feature vectors using an unsupervised clustering algorithm under each fault type group, classifying feature vectors with similar spatial distribution into the same feature cluster, and extracting the most representative k feature clusters of each fault type as its typical feature cluster;

[0099] Template standardization generation subunit: used to construct the average vector of the typical feature cluster of each fault type as the standardized template expression of this type of fault, and establish a one-to-one mapping structure between the fault type number and the feature template vector to form a fault template database for subsequent pattern projection.

[0100] Table 2 Fault template database:

[0101] ;

[0102] Table 2 above includes five common equipment faults. Three dominant independent feature dimensions are extracted for each fault, and the feature values ​​have been normalized. The fault type number is used to uniquely identify the fault type; the fault type name is named based on the labeled historical samples; the feature template vector dimensions 1-3 represent the normalized values ​​of the dominant feature components of the fault in the independent feature space; the dominant state parameters are the main influencing physical quantities identified based on the orthogonal feature-original parameter mapping results; the severity level reference provides a preset level benchmark for the subsequent level judgment unit, which can be dynamically updated in actual applications.

[0103] The above-mentioned sub-units cluster the representative feature vectors in known fault samples and refine them into standard templates, establish a clear "fault type-feature cluster" mapping relationship, and realize a templated and structured fault feature management mechanism. This provides a data basis and comparison standard for establishing an effective projection relationship between current working conditions and historical faults, enhancing recognition accuracy and system interpretability.

[0104] The fault level determination unit includes:

[0105] Amplitude change calculation subunit: used to perform difference analysis on the numerical amplitude of each feature dimension in the currently identified independent feature vector set and the corresponding fault template vector, obtain the feature offset, and calculate the fault offset coefficient based on the total amount of full-dimensional amplitude change , which is used to reflect the intensity characteristics of the fault manifestation. The formula is as follows: ,in, is the current eigenvector The value of the dimension, To match the corresponding dimension value in the fault template, is the characteristic dimension;

[0106] Time stability evaluation subunit: used to analyze the change trend of independent feature vectors in N consecutive time slices, and calculate the change fluctuation value using the mean square error indicator , which is used to express the time-continuous stability of the fault characteristics. The worse the stability, the more likely the fault is in a deteriorating state. The calculation formula is: ,in, For the The moment dimensional eigenvalues, is the average value at time N;

[0107] Spatial aggregation analysis subunit: used to determine the distribution concentration of the current fault characteristics in the sensor space coordinates, based on The distance between the centroid and the location of the fault impact point , used to represent the feature space focusing, and is calculated as: ,in, For the The spatial position vector of the fault point, is its spatial center of mass, The smaller the value, the higher the aggregation, which reflects that the fault area is clear;

[0108] Level determination output subunit: used to , , The three indicators are normalized and weighted to form a severity assessment score. The fault severity level (such as low, medium, and high) is output based on the scoring interval in the risk level standard table corresponding to the fault type. Finally, it is combined with the determined fault type code into structured data and embedded in the decoupled feature matrix under the current time index. By fusing the quantitative features of amplitude, time, and space dimensions and introducing risk level reference standards for different fault types, the unit realizes a multi-dimensional and accurate assessment of the severity of the current fault instance and outputs it in a structured manner, providing the system with a fault level result that is highly interpretable and has a clear decision-making basis, thereby improving the credibility and practical value of the diagnostic system.

[0109] Table 3 Risk level standards:

[0110] ;

[0111] In Table 3 above, the fault type number and fault type name are consistent with those in the aforementioned fault template database; F aThe reference range indicates the total amount of change in the current feature vector from the template vector; the σf reference range indicates the degree of feature fluctuation, and the larger the range, the more unstable the state; the Dc reference range indicates the distribution concentration of feature points in the sensor position; the severity level judgment rule represents the interval judgment logic combined with the three indicators, which is suitable for actual evaluation; the output level L is the level label finally generated by the system for structured data output, with values ​​of low / medium / high.

[0112] The interpretable diagnosis mapping module includes a feature-structure mapping unit, a fault location generation unit, and a life prediction output unit; wherein:

[0113] Feature-Structure Mapping Unit: This unit receives a decoupled feature matrix containing fault type codes and severity levels, extracts basic attributes such as structural component numbers, material categories, key component coordinates, and stress boundary conditions from the equipment physical parameter library, and establishes a one-to-one constraint mapping between feature dimensions and the equipment's physical structure based on the historical correspondence between fault types and dominant state parameters, identifying structural regions associated with current feature behaviors.

[0114] Fault location generation unit: This unit is used to extract the spatial location of the physical part corresponding to the current fault based on the mapping results. It then generates a structural map highlighting the current diagnosis results based on the sensor's 3D installation coordinates and the structural partition map. It then outputs the fault location map and embeds it into a visual diagnostic report in the form of structural number, component name, and coordinate position.

[0115] Life prediction output unit: Based on the evolution trajectory of the same type of faults in historical degradation data, combined with the current fault severity level and its time stability index, the remaining life statistical model of the corresponding component in the physical parameter library is retrieved, and the remaining life prediction value RUL (Remaining Useful Life) is output. It is presented as a quantitative indicator along with the fault location map in the final visual diagnosis report; the above unit constructs a feature-physical structure mapping channel and associates it with historical degradation patterns. This module realizes a two-way diagnostic transformation from abstract feature behavior to specific structural location and life assessment, giving the diagnostic report intuitive positioning and prediction capabilities, enhancing the system's industrial operability, technical interpretability and decision-making practicality.

[0116] The feature-structure mapping unit includes:

[0117] Parameter association analysis subunit: This subunit is used to extract the dominant state parameter set for each fault type from the established fault template database, query the device physical parameter library for structural parameter records that have a one-to-one response relationship with these state parameters, and construct an initial mapping set between feature dimensions and structural attributes.

[0118] Structural constraint matching subunit: used to introduce the amplitude weights of each feature dimension in the current decoupled feature matrix based on the initial mapping set and establish a constraint weight vector;

[0119] Regional correlation identification subunit: used to project the spatial coordinate information in the structural attribute set to the three-dimensional device model, aggregate high-response structural points based on the distribution density of the constraint weight vector on the spatial coordinates, and calculate the spatial response concentration ,when Exceeding the set threshold , then the region is determined to be the target related structure region; the calculation formula is: ,in, is the number of current response structure points, is the corresponding weight of the corresponding point, It is the spatial volume of the area where it is located or the volume of the cluster unit; by constructing a mapping path from the dominant characteristics to the structural parameters, and combining the current response weight and spatial distribution density for regional aggregation, this solution can accurately identify the key areas where the current fault characteristics are concentrated in the three-dimensional structural model of the equipment, realize high-precision diagnostic bridging from characteristic behavior to physical positioning, and enhance the practicality and explanatory ability of the system.

[0120] The present invention encompasses any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention. To provide a thorough understanding of the present invention, specific details are described in detail below in connection with the preferred embodiments of the present invention, but those skilled in the art will be able to fully understand the present invention without these detailed descriptions. Furthermore, to avoid unnecessary confusion regarding the essence of the present invention, well-known methods, processes, procedures, components, and circuits have not been described in detail.

[0121] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. Industrial equipment fault prediction and diagnosis system based on deep learning, characterized by: It includes a dynamic data alignment module, a multi-dimensional feature fusion module, a fault mode decoupling module, and an explainable diagnostic mapping module; among which: Dynamic data alignment module: used to align the multi-source heterogeneous sensor data of industrial equipment with temporal and spatial constraints, and build a multi-dimensional equipment status data stream under a unified timeline; Multi-dimensional feature fusion module: used to receive the aligned multi-dimensional device status data stream, perform orthogonal decomposition in the feature space, and extract a set of independent feature vectors; The multi-dimensional feature fusion module includes a state matrix construction unit, a feature correlation determination unit and an orthogonal feature extraction unit; wherein: State matrix construction unit: used to receive the unified time axis device state data stream output by the dynamic data alignment module, and align the spatial state of each sensor in each time slice. Converted into row vector form, stacked in time series to form the original state matrix, each column of the state matrix corresponds to a state parameter, and each row is a snapshot of the spatial state at a single time point; Feature correlation discrimination unit: used to perform correlation analysis on the state parameters of each column in the original state matrix. Through joint statistical indicators based on time and space dimensions, it identifies parameter pairs with multicollinearity or redundant coupling and marks their potential coupling structure. The feature correlation discrimination unit includes: Time correlation identification subunit: used to identify any two columns of state parameter vectors in the state matrix The Pearson correlation coefficient is calculated for the synergistic change trend on the entire time axis to identify parameter pairs with strong time correlation. The calculation formula is as follows: , where T is the length of the time series, and They are and If the mean of Greater than threshold , then it is determined to be a time-collinear parameter pair; Spatial synergistic change identification subunit: used to determine the spatial synergistic change relationship of multiple state parameters in a single time slice t, and calculate the parameters and Gather at sensor location The spatial structure similarity on , according to the number of each sensor , construct the following point pair covariance model: , where n is the total number of sensors, and Respectively represent The state value of a position, if Continuously greater than the space threshold , then it is marked as a spatial redundant pair; Redundant coupling marking subunit: It is used to perform intersection screening on the parameter pairs obtained through temporal correlation identification and spatial co-variation identification, extract the state parameter pairs that show strong coupling relationships in both temporal and spatial dimensions, and mark them as redundant coupling structures as feature compression objects in the subsequent orthogonal decomposition process; Orthogonal feature extraction unit: It is used to perform orthogonal decomposition processing based on feature correlation discrimination, reconstruct a set of independent feature vectors without correlation interference, and retain the physical meaning mapping relationship between the original parameters and the extracted features during the decoupling process to achieve structural restoration with physical interpretability; The orthogonal feature extraction unit comprises: Covariance matrix construction subunit: Constructs the covariance matrix C based on the state parameter set retained after feature correlation discrimination, which is used to characterize the linear relationship structure between each parameter; Feature orthogonalization subunit: used to perform eigenvalue decomposition on the covariance matrix C, extract its corresponding eigenvector set, and map the original state parameter set to the orthogonal feature space through projection transformation; Physical mapping preservation subunit: used to record the projection structure of each orthogonal eigenvector in the original state parameter space, calculate the feature contribution rate and reversely map the dominant physical parameters corresponding to each independent feature to form an orthogonal feature-original parameter mapping comparison table; Fault mode decoupling module: constructs a fault mode space projection template based on a set of independent feature vectors, and outputs a decoupled feature matrix containing fault type codes and fault severity levels; Explainable diagnostic mapping module: used to constrain matching of the decoupled feature matrix with the equipment physical parameter library to generate a visual diagnostic report containing a fault location map and remaining life prediction.

2. The industrial equipment fault prediction and diagnosis system based on deep learning according to claim 1 is characterized in that: The dynamic data alignment module includes a data acquisition unit, a time synchronization unit, a spatial registration unit, and a data interpolation unit; wherein: Data acquisition unit: used to synchronously collect original state parameter data from vibration sensors, temperature sensors, current and voltage sensors, and bearing acoustic emission sensors deployed on industrial equipment, including vibration acceleration values, surface temperature values, current values, voltage values, and acoustic emission signal amplitudes, and to attach original timestamps and 3D installation coordinates to each type of sensor data; Time synchronization unit: used to receive the state parameter data output by the data acquisition unit, align the timestamps of all state parameters using a sliding time window based on the set master time reference, and output a state data sequence with consistent time index; Spatial registration unit: It is used to receive the time-synchronized state data sequence, construct a spatial mapping function based on the three-dimensional installation coordinates of each sensor, and uniformly convert the position vectors of the state parameters to form a spatially consistent multi-dimensional device state data stream.

3. The deep learning-based industrial equipment fault prediction and diagnosis system according to claim 2, characterized in that: The spatial registration unit includes: Receive the three-dimensional installation coordinate information corresponding to each type of status parameter ,in Indicates the sensor number, Respectively represent The three-dimensional spatial position of each sensor in the industrial equipment coordinate system; Construct a local spatial reference frame R based on the structural geometry model of the industrial equipment, define the global alignment direction based on R, and transform all sensor coordinates to relative coordinates under the reference frame R. ; The aligned sensor coordinates Combined with the state parameter value at the corresponding moment, a unified format of spatial state pair is constructed. ; The state vectors of all sensors at the same time index By number Aggregate and output a spatially consistent multi-dimensional device status data stream, expressed as: , where n is the total number of sensors.

4. The industrial equipment fault prediction and diagnosis system based on deep learning according to claim 1 is characterized in that: The fault mode decoupling module includes a fault template construction unit, a mode projection execution unit and a fault level determination unit; wherein: Fault template construction unit: Based on the identified independent feature vector set and the characteristic response patterns of different types of faults in the historical fault sample set, typical fault feature clusters are extracted and organized into standardized fault pattern templates. The corresponding relationship between each fault type and its typical feature expression is established; Pattern projection execution unit: used to align and match the independent feature vector set extracted at the current moment or in the current cycle with the constructed fault pattern template. By calculating the similarity or distance index in the feature space, it determines which type of fault template the current feature distribution is closer to, and generates a fault type code with a unique identifier. Fault level discrimination unit: Based on the dimensions of amplitude change, temporal stability, and spatial aggregation of the feature vector, combined with the risk level standard corresponding to the fault type, the severity of the identified fault instance is evaluated, and the classification result is output. Structured data of fault type code + fault severity level is generated to form the final decoupled feature matrix.

5. The deep learning-based industrial equipment fault prediction and diagnosis system according to claim 4, characterized in that: The fault template construction unit includes: Feature label grouping subunit: used to group the data with labeled fault types in the historical fault sample set. Each group corresponds to a known fault type, and extracts the independent feature vector set within the corresponding time period to construct multiple sets of feature vector sequences indexed by fault type. Feature clustering analysis subunit: used to perform feature morphology clustering analysis on feature vectors using an unsupervised clustering algorithm under each fault type group, classifying feature vectors with similar spatial distribution into the same feature cluster, and extracting the most representative k feature clusters of each fault type as its typical feature cluster; Template standardization generation subunit: used to construct the average vector of the typical feature cluster of each fault type as the standardized template expression of this type of fault, and establish a one-to-one mapping structure between fault type number and feature template vector to form a fault template database.

6. The deep learning-based industrial equipment fault prediction and diagnosis system according to claim 4, characterized in that: The fault level determination unit includes: Amplitude change calculation subunit: used to perform difference analysis on the numerical amplitude of each feature dimension in the currently identified independent feature vector set and the corresponding fault template vector, obtain the feature offset, and calculate the fault offset coefficient based on the total amount of full-dimensional amplitude change , used to reflect the intensity characteristics of fault manifestation; Time stability evaluation subunit: used to analyze the change trend of independent feature vectors in N consecutive time slices, and calculate the change fluctuation value using the mean square error indicator , used to express the time-continuous stability of fault characteristics; Spatial aggregation analysis subunit: used to determine the distribution concentration of the current fault characteristics in the sensor space coordinates, based on The distance between the centroid and the location of the fault impact point , used to represent the feature space focality; Level determination output subunit: used to ,The three indicators are normalized and weighted ly combined to form a severity assessment score, and ,according to the scoring interval in the risk level standard table corresponding to the fault ,type, the fault severity level is outputted, and ,finally combined with the determined fault type coding into structured data, and ,embedded in the decoupled feature matrix under the current time index.

7. The deep learning-based industrial equipment fault prediction and diagnosis system according to claim 1, characterized in that: The explainable diagnostic mapping module includes a feature-structure mapping unit, a fault location generation unit, and a life prediction output unit; wherein: Feature-Structure Mapping Unit: This unit receives a decoupled feature matrix containing fault type codes and severity levels, extracts basic attributes such as structural component numbers, material categories, key component coordinates, and stress boundary conditions from the equipment physical parameter library, and establishes a one-to-one constraint mapping between feature dimensions and the equipment's physical structure based on the historical correspondence between fault types and dominant state parameters, identifying structural regions associated with current feature behaviors. Fault location generation unit: This unit is used to extract the spatial location of the physical part corresponding to the current fault based on the mapping results. It then generates a structural map highlighting the current diagnosis results based on the sensor's 3D installation coordinates and the structural partition map. It then outputs the fault location map and embeds it into a visual diagnostic report in the form of structural number, component name, and coordinate position. Life prediction output unit: Based on the evolution trajectory of the same type of faults in historical degradation data, combined with the current fault severity level and its time stability index, the remaining life statistical model of the corresponding component in the physical parameter library is retrieved, and the remaining life prediction value RUL is output. It is presented as a quantitative indicator together with the fault location map in the final visual diagnosis report.

8. The deep learning-based industrial equipment fault prediction and diagnosis system according to claim 7, characterized in that: The feature-structure mapping unit includes: Parameter association analysis subunit: This subunit is used to extract the dominant state parameter set for each fault type from the established fault template database, query the device physical parameter library for structural parameter records that have a one-to-one response relationship with these state parameters, and construct an initial mapping set between feature dimensions and structural attributes. Structural constraint matching subunit: used to introduce the amplitude weights of each feature dimension in the current decoupled feature matrix based on the initial mapping set and establish a constraint weight vector; Regional correlation identification subunit: used to project the spatial coordinate information in the structural attribute set to the three-dimensional device model, aggregate high-response structural points based on the distribution density of the constraint weight vector on the spatial coordinates, and calculate the spatial response concentration ,when Exceeding the set threshold , then the region is determined to be the target related structure region.

Images