Rolling bearing diagnosis system and method based on dynamic correction and neural symbolic reasoning

Abstract

The application discloses a rolling bearing diagnosis system and method based on dynamic correction and neural symbol reasoning. The system comprises a hardware shell for integrating a working condition sensing module, a neural symbol hybrid diagnosis engine, a storage-computation integrated edge execution unit and a general control module. The working condition sensing module is used for collecting bearing running state signals and mapping working condition characteristics. The neural symbol hybrid diagnosis engine is used for fusing physical rules and data characteristics for fault reasoning. The storage-computation integrated edge execution unit is used for accelerating the diagnosis process and guaranteeing offline running capability. The general control module is used for overall planning of timing scheduling, abnormality processing and result output. The above modules are electrically connected through the general control module, and their collaborative operation is controlled through a preset program. The application combines dynamic working condition online correction and neural symbol hybrid reasoning, realizes faster diagnosis and supports offline running, effectively balances interpretability and precision, strengthens few-sample generalization capability, and improves integration and facilitates deployment.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of bearing fault diagnosis, and in particular to a rolling bearing diagnosis system and method based on dynamic correction and neural symbolic reasoning. BACKGROUND

[0002] As a core component of rotating machinery, rolling bearings play a key role in high-reliability scenarios such as aircraft engines and wind power equipment, and their operating state is directly related to the safety and efficiency of the entire machine. Fault diagnosis technology, as an important means to ensure the reliable operation of rolling bearings, faces many challenges in complex industrial environments.

[0003] However, under dynamic working conditions, frequent fluctuations in equipment speed and load can cause nonlinear changes in bearing vibration characteristics, making it difficult for traditional diagnostic methods to capture these dynamic characteristics in real time, leading to misdiagnosis or missed diagnosis. In terms of hardware collaboration, existing diagnostic systems rely heavily on software algorithm optimization, which is not well adapted to edge computing devices, resulting in poor real-time performance and inability to maintain diagnostic functions in extreme situations such as network interruptions, which restricts the continuous monitoring needs of industrial sites. At the same time, current diagnostic techniques suffer from a disconnect between mechanism and data-driven approaches, making it difficult to meet the reliability requirements of diagnostic results in industrial scenarios. Therefore, developing a rolling bearing fault diagnosis system and method that can adapt to dynamic working conditions, deeply integrate hardware and algorithms, and balance the advantages of mechanism and data-driven approaches is of great significance for improving the health management level of industrial equipment and reducing operational costs. SUMMARY

[0004] The purpose of the present application is to provide a rolling bearing diagnosis system and method based on dynamic correction and neural symbolic reasoning, which can adapt to dynamic working condition changes, improve hardware collaboration efficiency, and achieve deep integration of mechanism and data-driven approaches to solve the problems raised in the background art.

[0005] To solve the above technical problems, the present application provides the following technical solution: a rolling bearing diagnosis system and method based on dynamic correction and neural symbolic reasoning, the system comprising: a hardware shell made of sheet metal process, the hardware shell is integrated with a working condition perception module, a neural symbol hybrid diagnosis engine, a storage-computing integrated edge execution unit and a total control module, the total control module is electrically connected with each of the above modules and controls their collaborative operation through a preset program, realizing the whole process diagnosis of rolling bearings from state perception to fault recognition.

[0006] Preferably, the working condition perception module comprises: a high-precision MEMS sensor array and a digital twin mapping unit, the MEMS sensor array adopts a distributed layout, integrates a rotating speed Hall sensor, an embedded temperature sensor and a three-axis vibration sensor, and realizes real-time acquisition of rotating speed, temperature and X / Y / Z three-axis vibration signals of the rolling bearing through a synchronous acquisition circuit; the digital twin mapping unit constructs an association model of bearing working condition parameters including rotating speed, load and temperature and fault features including impact frequency and kurtosis value based on Hertz contact theory, and maps the physical parameters into a feature vector that can be used for diagnosis.

[0007] Preferably, the operation method of the working condition perception module comprises the following steps:

[0008] S11: constructing an association model of working condition parameters and fault features;

[0009] S12: performing fault feature mapping;

[0010] S13: generating a feature vector.

[0011] Preferably, the neural-symbol hybrid diagnosis engine further comprises: a physical rule layer and a data-driven layer, the physical rule layer encodes the failure mechanism of the rolling bearing, including a fatigue crack propagation equation and a contact stress calculation model, and is used to dynamically constrain the output results of the data-driven layer; the data-driven layer comprises a lightweight graph convolution network (GCN) and an adaptive attention mechanism and an adaptive frequency band selection mechanism, the GCN is used to extract the spatio-temporal correlation features of multi-source signals including vibration, temperature and rotating speed, and the adaptive attention mechanism and the adaptive frequency band selection mechanism work cooperatively, learn and dynamically select the fault-sensitive frequency band, focus on key diagnostic information, and effectively improve the adaptability to variable working conditions. The neural-symbol hybrid diagnosis engine further integrates a fault causal chain modeling module, which uses causal reasoning technology to construct a directed acyclic graph of the failure mechanism described by the physical rule layer and the feature pattern identified by the data-driven layer, diagnose the fault type, reason the evolution path and root cause of the fault, and greatly enhance the explainability and decision support value of the diagnosis results.

[0012] Preferably, the neural-symbol hybrid diagnosis engine dynamically constrains the data-driven output through the physical rule, avoids misjudgment of a black box model, and the specific operation method is as follows:

[0013] S21: dynamic spatio-temporal feature extraction;

[0014] S22: wavelet packet decomposition of the vibration signal to extract sub-band energy;

[0015] S23: physical rule layer verification.

[0016] Preferably, the storage-computing integrated edge execution unit adopts a RISC-V architecture customized chip, integrates a 1D-CNN acceleration core and a wavelet transform special instruction set, and is used for hardware-level acceleration of feature extraction and inference process; the chip contains a near-memory computing design and a power consumption management module, and the SRAM in-memory computing unit can reduce data transfer power consumption and support independent diagnosis in an offline off-network state.

[0017] Preferably, the general control module is used for overall planning of timing scheduling, exception handling and result output.

[0018] Illustratively, the general control module adopts an STM32H743 microcontroller, is electrically connected with the working condition perception module, the neural-symbol hybrid diagnosis engine and the storage-computing integrated edge execution unit, is used for coordinating timing of the modules, processing exception feedback and outputting diagnosis results, and is further provided with an emergency stop switch for interrupting system operation in an emergency.

[0019] Compared with the prior art, the present application has the beneficial effects that: the present application combines dynamic working condition online correction and neural-symbol hybrid inference, can better adapt to variable working conditions, realizes faster diagnosis and supports offline operation, effectively balances explainability and precision, strengthens few-sample generalization ability, simultaneously improves integration degree and facilitates deployment. DETAILED DESCRIPTION

[0020] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, illustrate the present application and explain the technical scheme of the present application, and do not constitute a limitation on the present application. In the drawings:

[0021] Fig. 1 is a system structure schematic diagram provided by the embodiment of the present application;

[0022] Fig. 2 is a method step flow chart provided by the embodiment of the present application. DETAILED DESCRIPTION

[0023] The technical scheme in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0024] Embodiment 1

[0025] Please refer to Figs. 1-2The application provides the following technical scheme: a rolling bearing diagnosis system and method based on dynamic correction and neural symbol reasoning, the system comprising: a hardware shell made of sheet metal process, the hardware shell being integrated with a working condition sensing module, a neural symbol hybrid diagnosis engine, a memory-computing integrated edge execution unit and a general control module, the general control module being electrically connected with the above modules and controlling the cooperative operation of the modules through a preset program, realizing the whole-process diagnosis of the rolling bearing from state sensing to fault identification.

[0026] In the embodiment, the working condition sensing module is used for collecting bearing operation state signals and mapping working condition characteristics.

[0027] Illustratively, the working condition sensing module comprises: a high-precision MEMS sensor array and a digital twin mapping unit, the MEMS sensor array adopting a distributed layout, integrating a rotating speed Hall sensor (range 0-6000 rpm, accuracy ±0.2% FS), an embedded temperature sensor (range -40~125℃, accuracy ±0.5℃) and a three-axis vibration sensor (range ±16g, sampling rate 1-10kHz adjustable, noise density ≤20μg / √Hz), realizing the real-time collection of the rotating speed, temperature and X / Y / Z three-axis vibration signals of the rolling bearing through a synchronous collection circuit; the digital twin mapping unit is based on the Hertz contact theory, constructs an associated model of the bearing working condition parameters including rotating speed, load and temperature and the fault characteristics including impact frequency and kurtosis value, and maps the physical parameters into a feature vector that can be used for diagnosis.

[0028] In the embodiment, the working condition sensing module has the following operation method steps:

[0029] S11: constructing an associated model of working condition parameters and fault characteristics;

[0030] Illustratively, based on the Hertz contact theory, an explicit association between the bearing internal stress distribution and the fault characteristics is established, through dynamic load modeling, let the rotating speed be , the radial load be , the temperature be , and the thermal expansion coefficient be correct the bearing clearance:

[0031]

[0032] wherein, is the actual bearing clearance (unit: mm) at the temperature T, is the initial bearing clearance (unit: mm) at the reference temperature , and is the reference temperature (default 25℃, bearing design standard temperature).

[0033] Based on the corrected bearing clearance, the rolling element-raceway contact stress :

[0034] , It is the Poisson's ratio of bearing steel (typical value 0.3). It is the elastic modulus of bearing steel (typical value 210 GPa). This refers to the radius of curvature of the rolling elements. The bearing contact stress under dynamic load was quantified based on Hertzian contact theory. Its value increases with temperature ( (Increases) and decreases, with radial load The increase reflects the influence of the working conditions on mechanical stress.

[0035] S12: Perform fault feature mapping;

[0036] For example, generating the impact frequency and the outer ring fault characteristic frequency. :

[0037]

[0038] in, The bearing pitch circle diameter (the diameter of the center distribution circle of the rolling elements, in mm). The diameter of the rolling element (unit: mm). This is the contact angle (typically 15°). Bearing speed (unit: rpm, needs to be converted to Hz): , This indicates the number of rotations per second. This indicates the number of revolutions per minute. The formula is based on the rotational speed. and bearing geometry parameters ( The diameter of the pitch circle. The diameter of the rolling element, (where is the contact angle), calculate the characteristic frequency of outer ring faults, and use it to identify outer ring faults.

[0039] By introducing a load fluctuation factor Perform nonlinear kurtosis correction to correct the kurtosis of the vibration signal. The calculation formula is:

[0040]

[0041] in, The kurtosis of the original vibration signal is dimensionless. The exponential decay term of load fluctuation ( (This is the load fluctuation factor; the larger the value, the stronger the attenuation). The fatigue coefficient of the material is obtained through fatigue testing of bearing steel.

[0042] S13: generating a feature vector;

[0043] Exemplarily, the output includes a vector of outer ring fault characteristic frequency, vibration signal kurtosis, rolling element-raceway contact stress, and temperature Input the diagnostic engine.

[0044] In this embodiment, the neuro-symbolic hybrid diagnostic engine is used for fault reasoning by fusing physical rules and data features; the neuro-symbolic hybrid diagnostic engine further includes a physical rule layer and a data-driven layer, the physical rule layer encodes the failure mechanism of the rolling bearing, including a fatigue crack propagation equation, a contact stress calculation model, etc., and is used to dynamically constrain the output result of the data-driven layer; the data-driven layer includes a lightweight graph convolutional network (GCN), an adaptive attention mechanism, and an adaptive frequency band selection mechanism, the GCN is used to extract the spatio-temporal correlation features of the multi-source signals including vibration, temperature, and rotating speed, and the adaptive attention mechanism and the adaptive frequency band selection mechanism work cooperatively: first, the adaptive frequency band selection mechanism dynamically selects candidate frequency bands with energy mutation or significant changes in entropy value based on the power spectral density of real-time signals; then, the adaptive attention mechanism learns and gives higher weights to the frequency bands related to the outer ring, inner ring, etc. fault characteristic frequency based on the candidate frequency bands and the historical fault database (stored in the edge execution unit), so as to realize accurate focusing on key diagnostic information.

[0045] In this embodiment, the neuro-symbolic hybrid diagnostic engine dynamically constrains the data-driven output through physical rules to avoid misjudgment of black box models, and the specific operation method is as follows:

[0046] S21: dynamic spatio-temporal feature extraction;

[0047] Exemplarily, a dynamic adjacency matrix is constructed based on the physical position relationship of the sensors , , is an adjacency matrix, is an identity matrix (the purpose of introducing the identity matrix is to add a "self-loop" to the adjacency matrix, that is, each sensor node is connected with itself to ensure that the feature information of the node itself is retained in the graph convolutional network (GCN) operation), then the Euclidean distance weight is calculated by calculating the spatial geometric correlation of the three-axis vibration sensor, and the mutual information weight is calculated by calculating the time sequence correlation of the temperature / rotating speed sensor, and the calculation formula is:

[0048]

[0049] wherein, is the adjacency weight between sensor i and sensor j (dimensionless, reflecting the correlation strength between the two), is the sensor position coordinate, is the time sequence weight coefficient, Mutual information, time series signal collected by sensor i (such as vibration, temperature, rotation speed, etc.), time series signal collected by sensor j; Mutual information threshold (when two sensors are considered to be related in time series). This formula constructs the adjacency weight between sensors, combining the spatial Euclidean distance (Dij) , sensor coordinates) and the time series mutual information (ωij is the time series weight coefficient), reflecting the spatial and temporal correlation of the sensors.

[0050] For example, the data-driven layer uses a lightweight GCN to input multi-sensor time series data , three-axis vibration + temperature + rotation speed, is a real set, indicating that the data is real, is the time step, indicating that the collected time series data contains t samples at t time points), input : actual time series data containing three-axis vibration, temperature, and rotation speed, for spatio-temporal feature extraction:

[0051] where, is a linear rectifier activation function used to introduce nonlinear features, l is the number of layers of the graph convolution network (the lth layer), is the degree matrix (diagonal matrix, diagonal elements are the degrees of each sensor node, i.e. the number of connections with other nodes), is the feature matrix (output features of the lth layer, is the output feature of the lth layer), is the adjacency matrix with self-loop, is the learnable weight.

[0052] S22: Wavelet packet decomposition of the vibration signal to extract sub-band energy, i.e. using the sub-band energy of wavelet packet decomposition :

[0053]

[0054] where, is the discrete signal of the ith sub-band (k = 1, 2, 3,..., N), and N is the signal length.

[0055] Calculate the frequency band weight :

[0056] where, is the sensitivity coefficient, the average energy of the historical failure frequency band, the energy of the jth frequency band, and the same dimension, used for normalization calculation). This formula calculates the weight of each frequency band, focusing on the key frequency band related to failure through the difference from the average energy of the historical failure frequency band.

[0057] S23: Perform physical rule layer verification;

[0058] For example, using fatigue crack propagation verification, if the GCN output failure type is "fatigue peeling", the following conditions need to be met:

[0059]

[0060] where, is the crack length increment (mm), is the cycle increment (times), is the material fatigue index (dimensionless, determined by material properties), is the stress intensity factor amplitude, is the fatigue crack propagation threshold value (Kth) , when crack does not expand); if not met, the result is rejected, and the physical rule layer verification is output and corrected. By verifying the physical reasonableness of the "fatigue peeling" failure, only when the crack propagation rate meets the material properties ( , material constant) and the stress intensity factor amplitude exceeds the threshold value , the diagnosis result is valid.

[0061] Output correction: force the result that does not meet the mechanism into the review process, triggering sensor resampling.

[0062] In addition, the neural symbol hybrid diagnosis engine of the embodiment also integrates a failure cause chain modeling module. This module is started after the physical rule layer verification (S23), and when a specific failure (such as outer ring peeling) is diagnosed, the module is activated. Based on the preset causal network, it analyzes the root cause that may lead to the failure, for example: whether the load in the current and historical data is continuously over-standard, whether the lubrication state (indirectly reflected by temperature) is abnormal, etc., and outputs "high probability root cause: long-term overloading operation" or "suspected poor lubrication" and other diagnostic conclusions with causal relationship, providing deeper insights for operation and maintenance decisions.

[0063] In the embodiment, the storage-computing integrated edge execution unit is used to accelerate the diagnosis process and guarantee offline operation capability;

[0064] ​Exemplarily, the storage-computing integrated edge execution unit adopts a RISC-V architecture customized chip, integrates a 1D-CNN acceleration core and a wavelet transform special instruction set, and is used for hardware-level acceleration of feature extraction and inference process; the chip contains a near-memory computing design and a power consumption management module, and the SRAM in-memory computing unit can reduce data transfer power consumption and support independent diagnosis in an offline off-network state.

[0065] In the embodiment, the total control module is used for overall planning of timing scheduling, exception handling and result output.

[0066] Exemplarily, the total control module adopts an STM32H743 microcontroller, is electrically connected with the working condition perception module, the neural-symbol hybrid diagnosis engine and the storage-computing integrated edge execution unit, is used for coordinating timing of the modules, processing exception feedback and outputting a diagnosis result, and is further provided with an emergency stop switch for interrupting system operation in an emergency.

[0067] It should be noted that, in this document, the relationship terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device.

[0068] Finally, it should be noted that: the above only describes the preferred embodiments of the present application, and is not used to limit the present application, although the present application has been described in detail with reference to the foregoing embodiments, and for those skilled in the art, the technical solutions recorded in the foregoing embodiments can still be modified, or some technical features can be replaced equivalently. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A rolling bearing diagnostic system based on dynamic correction and neuro-symbolic reasoning, characterized in that: The application relates to a hardware shell made of sheet metal, which is internally integrated with a working condition sensing module, a neural-symbol hybrid diagnosis engine, a storage-computing integrated edge execution unit and a general control module, the working condition sensing module is used for collecting bearing operation state signals and mapping working condition characteristics, the neural-symbol hybrid diagnosis engine is used for fusing physical rules and data characteristics to perform fault reasoning, the storage-computing integrated edge execution unit is used for accelerating the diagnosis process and guaranteeing offline operation capability, and the general control module is used for overall planning of time sequence scheduling, abnormality processing and result output, the above modules are electrically connected through the general control module, and the modules are controlled to cooperatively operate through a preset program, so that full-process diagnosis from state sensing to fault identification of a rolling bearing is realized. The working condition sensing module comprises a high-precision MEMS sensor array and a digital twin mapping unit; the high-precision MEMS sensor array adopts a distributed layout and is integrated with a rotating speed Hall sensor, an embedded temperature sensor and three-axis vibration sensors, realizes real-time acquisition of rotating speed, temperature and X / Y / Z three-axis vibration signals of the rolling bearing through a synchronous acquisition circuit, and maps physical parameters into feature vectors which can be used for diagnosis based on a Hertz contact theory. The neural-symbol hybrid diagnosis engine further comprises a physical rule layer and a data driven layer; the physical rule layer is based on encoding rolling bearing failure mechanism and comprises a fatigue crack propagation equation and a contact stress calculation model, and is used for dynamically constraining output results of the data driven layer; the data driven layer comprises a lightweight graph convolution network (GCN) and an adaptive attention mechanism; the GCN is used for extracting space-time correlation features of multi-source signals including vibration, temperature and rotating speed; and the adaptive attention mechanism focuses on key diagnosis information by learning a fault sensitive frequency band. The storage-computing integrated edge execution unit adopts a RISC-V architecture customized chip, integrates a 1D-CNN acceleration core and a wavelet transform special instruction set, is used for hardware-level acceleration of feature extraction and reasoning process, and comprises a near memory computing design and a power consumption management module; an SRAM in-memory computing unit of the chip can reduce data transfer power consumption and support independent diagnosis in an offline off-network state.

2. The dynamic correction and neural-symbolic reasoning based rolling bearing diagnosis system according to claim 1, performing a dynamic correction and neural-symbolic reasoning based rolling bearing diagnosis method, characterized in that: The method comprises a working condition sensing module operation method, and steps are as follows: S11: an associated model of working condition parameters and fault characteristics is constructed; S12: fault characteristic mapping is performed; S13: a feature vector is generated.

3. The dynamic correction and neural-symbolic inference based rolling element bearing diagnostic system of claim 2, wherein: The associated model of working condition parameters and fault characteristics is constructed by establishing an explicit association between bearing internal stress distribution and fault characteristics based on the Hertz contact theory, performing dynamic load modeling, correcting bearing clearance through a thermal expansion coefficient, and further calculating rolling element-raceway contact stress based on the corrected bearing clearance.

4. The dynamic correction and neural-symbolic reasoning based rolling element bearing diagnostic system of claim 3, wherein: In the neural-symbol hybrid diagnosis engine, the data driven output is dynamically constrained by physical rules to avoid misjudgment of a black box model, and specific operation methods are as follows: S21: dynamic space-time feature extraction is performed; S22: wavelet packet decomposition is performed on the vibration signal to extract sub-band energy; S23: physical rule layer verification is performed.

5. The dynamic correction and neural-symbolic reasoning based rolling element bearing diagnostic system of claim 4, wherein: In dynamic spatio-temporal feature extraction, the spatial and temporal correlation of sensors is obtained by constructing the adjacency weight between sensors, combining the spatial Euclidean distance and the temporal mutual information, and then using a lightweight GCN based on the data-driven layer to input the multi-sensor time series data including three-axis vibration, temperature and speed for spatio-temporal feature extraction.

6. The dynamic correction and neural-symbolic reasoning based rolling bearing diagnostic system of claim 5, wherein: The total control module adopts an STM32H743 microcontroller, is electrically connected with the working condition perception module, the neural-symbol hybrid diagnosis engine and the storage-computation integrated edge execution unit, coordinates the time sequence of each module, processes abnormal feedback and outputs a diagnosis result, and is further provided with an emergency stop switch for interrupting system operation in an emergency.

7. The dynamic correction and neural-symbolic reasoning based rolling bearing diagnostic system of claim 6, wherein: The fault-sensitive frequency band of the adaptive attention mechanism not only includes the bearing outer ring fault characteristic frequency and the bearing inner ring fault characteristic frequency, but also dynamically adjusts the attention weight by calculating the difference between the energy of each frequency band and the historical fault mode, and strengthens the extraction of key diagnostic information. The data-driven layer further includes an adaptive frequency band selection mechanism which dynamically selects the frequency band with the most diagnostic value based on the real-time features of the power spectral density or entropy value of the input signal, thereby improving the sensitivity of fault diagnosis and the adaptability to variable working conditions. The neural-symbol hybrid diagnosis engine further integrates a fault causal chain modeling module which uses a causal network modeling method to construct a rolling bearing fault evolution path, associates the failure mechanism in the physical rule layer with the feature mode identified by the data-driven layer, and reasons the existing fault root cause.

Images