An event camera based bearing cross-speed fault diagnosis method and system driven by angular domain phase representation

Abstract

The application discloses a bearing cross-speed fault diagnosis method and system based on an event camera angular domain phase representation drive, first determines the synchronous public interval of multi-view original event flow and carries out cutting, then carries out fixed rotation number trigger type window cutting to obtain samples, constructs double-channel event phase sequences based on the total event intensity and multi-direction differential response intensity of the event fragments of the samples, and then carries out angular domain resampling and multi-view adaptive fusion based on impact significance, extracts the impact statistical prior of the fused double-channel angular domain phase sequences, generates a conditional gating vector according to the impact statistical prior, carries out channel-by-channel modulation on the angular domain long sequence feature transmission of a residual Mamba module, and further realizes bearing cross-speed fault recognition. The technology makes the event phase construction, angular domain statistical prior and state space feature extraction form a continuous correlation, improves the physical phase comparability of different speed samples and the selective representation ability of the diagnostic model for impact related features.

Description

Technical Field

[0001] This invention belongs to the field of bearing fault diagnosis technology, and in particular relates to a bearing cross-speed fault diagnosis method and system based on angular domain phase characterization driven by an event camera. Background Technology

[0002] Rolling bearings are critical components in rotating machinery systems, and their health directly affects equipment operational safety, maintenance costs, and service life. Existing bearing fault diagnosis solutions typically rely on contact-based acquisition methods such as accelerometers, acoustic emission sensors, and current sensors to obtain vibration or condition information. While these solutions are widely used, they still suffer from inconvenient installation, complex maintenance, and insufficient adaptability to operating conditions, such as high-speed rotation, limited installation space, complex on-site environments, difficulties in deploying contact sensors, or the inability to introduce additional loads.

[0003] Event cameras are a type of visual sensor that outputs asynchronous events when pixel brightness changes exceed a preset threshold. They feature microsecond-level temporal resolution, high dynamic range, and low redundancy output. Compared to traditional frame cameras, event cameras are better suited for capturing localized impacts, weak vibrations, and high-speed dynamic responses in rotating machinery, thus providing a new path for non-contact condition monitoring of mechanical equipment.

[0004] Existing technologies already include non-contact mechanical vibration monitoring and fault diagnosis schemes based on event cameras. For example, CN116734980A discloses a non-contact mechanical vibration monitoring and fault diagnosis method based on an event camera. This method uses an event camera to collect vibration event signals of rotating machinery, converts the event stream into a continuous sequence of event frames, and extracts fault feature frequencies through Gabor filters, wavelet packet decomposition, and envelope spectrum analysis, thereby achieving mechanical vibration monitoring and fault diagnosis. This scheme can achieve non-contact vibration sensing using an event camera, but it mainly extracts vibration time-domain signals and frequency-domain features in the time domain. When the rotational speed of the bearing under test changes, the response of similar faults will be compressed or stretched on the time axis, making it difficult to maintain the same mechanical angular position correspondence between samples at different rotational speeds. In addition, existing rolling bearing diagnosis schemes based on event cameras usually reconstruct the event stream into a two-dimensional event representation or event frames and then feed it into a neural network for classification. This type of method can verify the feasibility of using event cameras for bearing fault diagnosis, but it usually does not form a complete processing chain to address the problems of mechanical phase alignment under cross-rotational speed conditions, the reliability difference of response from multiple observation perspectives, and the instability of feature scale caused by event density fluctuations.

[0005] Therefore, existing technologies face at least the following technical obstacles: First, fixed-length or time-domain event frame sequences cannot guarantee that samples at different rotational speeds cover the same mechanical phase range; second, directly using event counts or event frame intensity is easily affected by event density fluctuations, sparse responses, and local noise; third, under multi-view event acquisition, different views have different sensitivities to fault impact, and simple averaging or splicing easily introduces low-quality view noise; fourth, existing diagnostic models typically directly splice manual statistical features with deep temporal features during the classification stage, or only use the preprocessing results as input to a regular network, failing to utilize the impact peak intensity, transient amplitude, and phase distribution asymmetry reflected by the event angular domain phase sequence to conditionally adjust the long-sequence feature transmission process within the network, resulting in insufficient interaction between front-end event feature construction and back-end network feature extraction. Therefore, further exploration of fault diagnosis technologies based on event cameras is needed. Summary of the Invention

[0006] The purpose of this invention is to provide a bearing cross-speed fault diagnosis method and system based on angular domain phase representation driven by event cameras, in order to solve the problems in the prior art where it is difficult to directly align and compare similar fault event streams at different speeds, and the fusion effect is unstable due to inconsistent multi-view response quality, especially the decrease in diagnostic accuracy in cross-speed scenarios. This invention can transform multi-view asynchronous event streams into angular domain features with comparable physical phases, and further utilizes the unified technical solution of the impact statistical prior modulation of angular domain features to modulate the internal feature transfer process of the diagnostic network.

[0007] Therefore, the present invention provides the following technical solution:

[0008] A bearing cross-speed fault diagnosis method based on angular domain phase characterization driven by an event camera includes the following steps.

[0009] Step 1: Multi-view event stream acquisition and synchronization, that is, acquiring event stream data of the bearing under test from at least two observation perspectives through the event camera and clipping the event streams from different observation perspectives to a common time interval.

[0010] Among them, by pruning the common time interval, only events with timestamps located within the common time interval are retained, and redundant event data at the front or back end that do not belong to the common time interval are removed from each observation perspective, so as to ensure that the event segments extracted from different observation perspectives correspond to the same mechanical operation process.

[0011] Step 2: Sample Construction. Event segments are extracted from the event stream according to the window length of a fixed revolutions window as event segment samples. The fixed revolutions window is designed based on the bearing speed and meets the requirement of covering the mechanical revolutions. Fixed dynamic duration window;

[0012] Step 3: Constructing a dual-channel event phase sequence. For each event segment sample, time slices are divided, and event frames are constructed by accumulating events in the same time slice. The total event intensity and differential response intensity in at least two directions are then calculated. The dual-channel event phase sequence is constructed using the total event intensity and differential response intensity.

[0013] Step 4: Corner domain resampling and multi-view adaptive fusion. The dual-channel event phase sequence corresponding to each event fragment sample under each observation view is resampled to a uniform corner domain length to obtain a dual-channel corner domain phase sequence; then the dual-channel corner domain phase sequences of all observation views are fused to obtain the dual-channel corner domain input after event fragment sample fusion.

[0014] Step 5: Send the dual-channel angular domain input of the event fragment sample to the fault diagnosis model to obtain the fault classification probability of the event fragment sample, and then combine the fault classification probabilities of the event fragment samples on the same event stream to obtain the bearing fault category.

[0015] Optionally, the process of constructing a dual-channel event phase sequence using the total event intensity and the differential response intensity in step 3 can be represented by the following mathematical model:

[0016] ;

[0017] ;

[0018] ;

[0019] ;

[0020] In the formula, Indicates the first The observation perspective in the first The total event intensity of each time slice is used to reflect the overall level of event activity within the time slice; Indicates the first The observation perspective in the first A type of differential intensity for each time slice; Indicates the first The observation perspective in the first Another type of differential intensity for a time slice; and They represent the first The observation perspective in the first One type of event phase and another type of event phase in a time slice; and All of these are minimal constants used to prevent the denominator from being zero; Indicates the first The number of time slices obtained by dividing the current fixed revolutions window from each observation perspective; and This represents the mean-free dual-channel event phase sequence, corresponding to one type of event phase and another type of event phase; For the first The adaptive scaling factor corresponding to the current fixed revolutions window for each observation view; It is the arctangent function in the four quadrants.

[0021] Optionally, adaptive scaling factor An adaptive scaling factor based on median absolute deviation is used, and each fixed-rotation window under each observation view corresponds to one adaptive scaling factor. Specifically:

[0022] ;

[0023] ;

[0024] ;

[0025] in, Indicates the absolute deviation of the median. and All are minimal constants to prevent the denominator from being zero, and z is a user-defined parameter symbol; Indicates the first The observation perspective in the first The combined differential response consisting of two types of differential intensities in each time slice; Indicates the first The composite differential response sequence consisting of all time slices from each observation perspective; Indicates the first The total event intensity sequence consisting of all time slices from each observation perspective; This indicates that the value is restricted to a certain range. Truncation operator within range, and These are the minimum and maximum values ​​of the preset scaling factor, respectively.

[0026] Optionally, in step S3, each event segment sample is divided into time slices, and then an event frame is constructed by accumulating events within the same time slice. An event frame is defined as follows:

[0027] ;

[0028] in, Indicates the first The observation perspective in the first Pixels in an event frame formed by a time slice The cumulative event value at the location; Indicates the first A sample of an event fragment from a single observation perspective; Indicates the first The event polarity of an event; This is an indicator function that takes the value 1 if the condition within the parentheses is true, and 0 otherwise. Indicates the first A time slice;

[0029] For the i-th event in the v-th observation view, the indicator function is valid only if the following conditions are met simultaneously. Choose 1:

[0030] (1) The event belongs to the current event fragment sample. ;

[0031] (2) Event timestamp Located in the nth time slice Inside;

[0032] (3) The x-axis of the event It equals the x-coordinate of the current pixel;

[0033] (4) The vertical axis of the event It equals the y-coordinate of the current pixel.

[0034] Optionally, the formula for calculating the fixed revolutions window in step 2 is:

[0035] ;

[0036] in, This indicates the bearing speed, measured in revolutions per minute (rpm). Indicates a single revolution cycle, in seconds; This indicates the duration of the event segment extracted from the event stream, i.e., the window duration of a fixed-revolution window;

[0037] The process of generating the event fragment samples is as follows: within the common time interval, the event counts of each observation perspective are counted with a bin step size to obtain the event energy in each time bin, thereby forming a total energy sequence; sampling anchor points are selected on the total energy sequence, and the time of the sampling anchor point is taken as the starting time. Event fragments are extracted from the event stream according to the window length of a fixed number of revolutions window to construct the event fragment samples.

[0038] Optionally, in step 4, the viewpoint fusion weight is determined based on the impact significance of the dual-channel angular domain phase sequences corresponding to each observation viewpoint, and then the dual-channel angular domain phase sequences of all observation viewpoints are fused using the viewpoint fusion weight, specifically as follows:

[0039] First, construct intermediate representations:

[0040] ;

[0041] Then, Pearson kurtosis is calculated, and the Pearson kurtosis of each observation view is transformed into initial view fusion weights using Softmax, and then converted into final view fusion weights:

[0042] ;

[0043] ;

[0044] in, and They represent the first The observation perspective in the first The two angular domain phase channels at each angular domain sampling point are obtained by resampling the dual-channel event phase sequence corresponding to each observation view to a uniform angular domain length. For the first Intermediate representation of current event fragment samples from various observation perspectives; , , respectively, are the Pearson kurtosis corresponding to the v-th and r-th observation viewpoints, used to characterize the impact significance in the angular domain phase sequence; These are the control parameters for the weight distribution; Weigh the initial perspectives; For smoothing coefficients; Weights are integrated for the final perspective;

[0045] The fusion model is as follows:

[0046] ;

[0047] ;

[0048] in, , These are the two channel variables of the fused dual-channel angular domain input X, with dimensions of... ; This represents the number of observation angles.

[0049] Optionally, the fault diagnosis model includes a angular domain temporal backbone branch, a statistical prior generation branch, a conditional gating module, and a classification head; the dual-channel angular domain input is... The input angular domain temporal backbone branch and the statistical prior generation branch are respectively used. The statistical prior generation branch utilizes the fused dual-channel angular domain input of the current event segment samples. Generate event-angle domain statistical prior vectors The conditional gating module will statistically analyze the prior vectors of the event angle domain. Mapped to a conditional gating vector;

[0050] The angular domain time-series backbone branch has at least a plurality of sequentially connected residual Mamba modules. For each residual Mamba module, the angular domain long sequence features output by the residual Mamba module are modulated channel by channel using the conditional gating vector.

[0051] The modulated features corresponding to the last residual Mamba module are processed by the classification head to output the fault classification probability of each event segment sample. The fault classification result is then obtained by combining the fault classification probabilities of event segment samples from the same event stream. This invention also provides an angular domain phase representation-driven event camera bearing cross-speed fault diagnosis system, comprising the following modules:

[0052] The multi-view event stream acquisition and synchronization module is used to acquire event stream data of the bearing under test from at least two observation views through the event camera and to perform common time interval clipping on the event streams from different observation views.

[0053] The sample construction module is used to extract event fragments from the event stream as event fragment samples according to the window length of a fixed revolutions window. The fixed revolutions window is designed based on the bearing speed and meets the requirement of covering the mechanical revolutions. Fixed dynamic duration window;

[0054] The dual-channel event phase sequence construction module is used to divide each event segment sample into time slices, and then construct event frames by accumulating events in the same time slice. The total event intensity and differential response intensity in at least two directions are then calculated, and the dual-channel event phase sequence is constructed using the total event intensity and differential response intensity.

[0055] The corner domain resampling and multi-view adaptive fusion module is used to resample the dual-channel event phase sequence corresponding to each event segment sample under each observation view to a unified corner domain length to obtain a dual-channel corner domain phase sequence; then, the dual-channel corner domain phase sequences of all observation views are fused to obtain the dual-channel corner domain input after the event segment sample fusion.

[0056] The diagnostic output module is used to send the dual-channel angular domain input of the event fragment samples to the fault diagnosis model to obtain the fault classification probability of the event fragment samples, and then combine the fault classification probabilities of the event fragment samples on the same event stream to obtain the bearing fault category.

[0057] The present invention also provides a computer device, comprising: one or more processors and a memory storing one or more computer programs;

[0058] The processor invokes the computer program to implement the steps of a bearing cross-speed fault diagnosis method driven by angular domain phase characterization based on an event camera.

[0059] The present invention also provides a computer-readable storage medium storing a computer program that is invoked by a processor to implement the steps of a bearing cross-speed fault diagnosis method driven by angular domain phase characterization based on an event camera.

[0060] Compared with existing methods, the present invention achieves the following progress and effects:

[0061] 1. This invention uses common time interval clipping, fixed rotation window, and angular domain resampling to uniformly map event segments with different durations at different rotation speeds to the same mechanical angular domain coordinate system, enabling samples of different rotation speeds to be compared within the same rotation range and at corresponding angular positions. This overcomes the problem that time domain event sequences are difficult to align directly due to stretching or compression caused by changes in rotation speed, and effectively solves the technical obstacles existing in the prior art.

[0062] 2. In the preferred embodiment provided by this invention, a total event intensity and two types of differential response intensities are constructed based on event fragments, and a bounded dual-channel event phase sequence is formed through arctangent phase mapping. Specifically, the total event activity level is used as a reference quantity, and the event structure changes in two directions are used as response quantities. Then, MAD is used to perform window-level adaptive matching of the scales of the two, ultimately forming a bounded and relatively scale-stable dual-channel event phase sequence. This processing method can simultaneously retain the overall event activity level and directional structure change information, reducing the feature scale instability caused by changes in the number of events, event density, and local noise of different samples.

[0063] 3. In the preferred embodiment provided by this invention, the dual-channel event phase sequence is resampled to a uniform angular domain length, allowing samples at different rotational speeds to be compared at the same mechanical angular position. The fusion weight is determined based on the impact significance of the angular domain phase sequence at each observation perspective. Compared to simple averaging, fixed weighting, or direct splicing, this invention can improve the contribution of perspectives with significant impact responses and suppress the interference of weak responses or approximately constant perspectives on the fusion results, thereby overcoming the problem of decreased diagnostic stability across rotational speeds due to inconsistent multi-view response quality.

[0064] 4. In the preferred embodiment provided by this invention, kurtosis, peak value, and skewness are extracted from the fused dual-channel angular domain phase sequence as statistical priors for the event angular domain. Conditional gating vectors are then generated from these statistical priors to modulate the state-space features output by the residual Mamba module channel by channel. Compared to simply concatenating statistical features before the classification head, this invention allows the impact spike intensity, transient response strength, and phase distribution asymmetry of the event angular domain phase to directly participate in the long-sequence feature transmission process within the network. This establishes a continuous interaction between the front-end event phase construction and the back-end state-space modeling, which is beneficial for enhancing the model's selective representation ability of impact-related angular domain features. Attached Figure Description

[0065] Figure 1 This is a schematic diagram of the basic process of the method in an embodiment of the present invention.

[0066] Figure 2 This is a diagram illustrating a specific diagnostic process in an embodiment of the present invention.

[0067] Figure 3 This is a schematic diagram of the differential enhancement event phase feature construction process in an embodiment of the present invention.

[0068] Figure 4 This is a flowchart of the corner domain resampling and multi-view adaptive fusion process in an embodiment of the present invention. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. The technical features involved in the various embodiments of the invention described below can be combined with each other as long as they do not conflict with each other.

[0070] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0071] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0072] This invention provides a bearing cross-speed fault diagnosis method and system based on angular domain phase representation driven by an event camera, applicable to fault diagnosis of motor bearings. Its core improvements include: firstly, by using common time interval pruning, a fixed rotational speed window, and angular domain resampling, event segments with different durations at different speeds are uniformly mapped to the same mechanical angular domain coordinate system; secondly, by constructing the total event intensity and differential response intensity in at least two directions, and combining median absolute deviation adaptive scaling, unified angular domain mapping, and multi-view impact saliency fusion to obtain dual-channel angular domain input; further, event angular domain statistical priors are extracted from the dual-channel angular domain input, and conditional gating vectors are generated based on these priors. Then, the angular domain long-sequence feature transfer process of the residual Mamba module is modulated using the conditional gating vectors, achieving collaborative utilization of the front-end event phase representation and the back-end state-space diagnostic network for fault diagnosis.

[0073] The present invention will be further described below with reference to embodiments.

[0074] Example 1:

[0075] like Figure 1 and Figure 2 As shown, this embodiment provides a bearing cross-speed fault diagnosis method based on angular domain phase characterization driven by an event camera, including the following steps:

[0076] Step 1: Acquisition and synchronization of multi-view event streams.

[0077] Step 1-1: Event Stream Acquisition; In this embodiment of the invention, multiple event cameras are arranged around the area of ​​the bearing under test. The placement of the event cameras is not limited to a fixed absolute angle, but should meet the following conditions: the field of view of each event camera covers the observable area of ​​the bearing under test or its adjacent rotating structure; each event camera remains relatively fixed to the device under test during the acquisition process; there are at least two observation angles with non-completely overlapping observation directions to obtain vibration or impact responses under different projection directions. Preferably, three event cameras are used to simultaneously observe the area of ​​the bearing under test from different directions to reduce information loss caused by occlusion, insufficient local texture, or insensitivity to projection direction from a single perspective.

[0078] Taking three event cameras as an example, the raw event stream data output by each event camera... It can be represented as:

[0079] ;

[0080] Where v represents the observation viewpoint number, Represents a timestamp. For pixel coordinates, Indicates the polarity of the event. This represents the number of events corresponding to the v-th observation viewpoint. The event refers to the asynchronous sampling result output when the brightness change at a certain pixel position in the event camera reaches the preset trigger condition. i is the event marker. This represents the event stream data corresponding to the i-th event. For pixel coordinates, a positive polarity event is output when the change in logarithmic brightness relative to the reference time is greater than a positive threshold; a negative polarity event is output when the change in logarithmic brightness is less than a negative threshold. Event polarity is represented as: ,in, This indicates an event triggered by an increase in brightness at that pixel. This indicates an event triggered by a decrease in brightness at that pixel. It should be understood that, for rotating machinery scenes, brightness changes mainly originate from local positional and contrast changes in the image plane of the surface texture, edges, and marked areas of the bearing or shaft under test due to vibration, impact, or rotational motion.

[0081] Steps 1-2: Common time interval pruning of the event stream; after decoding the original event stream data from each observation perspective, first determine if there is any out-of-order issue with the timestamps; if so, rearrange the event stream according to the timestamps; then obtain the common time interval of the event stream from each observation perspective:

[0082] ;

[0083] In the formula, The common time interval is defined as the start and end times of the common time zone, unifying the common time interval for all observation perspectives. Specifically, it is the intersection of the event flow time ranges of each observation perspective, i.e., the maximum value among the start times of each perspective is taken as the common start time, and the minimum value among the end times of each perspective is taken as the common end time. Subsequent observation perspectives will only retain events within this common time interval to ensure that event segments from multiple perspectives correspond to the same mechanical operation process. These refer to the first event from the v-th observation perspective and the... The timestamps of each event are represented by V, which indicates the number of observation perspectives. In this embodiment, the common time interval of the event streams from three observation perspectives is taken, and the remaining redundant time intervals are removed. That is, after obtaining the common time interval, only events whose timestamps are located within the common time interval are retained, and redundant event data at the beginning or end of each observation perspective that do not belong to the common time interval are removed, so as to ensure that the event segments subsequently extracted from different observation perspectives correspond to the same mechanical operation process.

[0084] Step 2: Constructing event fragment samples.

[0085] Step 2-1: Construct the total energy sequence. Within a common time interval, count the events from each observation perspective using a binning step size to obtain the event energy within each time bin. For example, the preset binning step size is 1 millisecond; it should be understood that this value is empirical and not specifically limited in this invention. The total energy sequence is obtained by superimposing the event counts from each observation perspective within the same time bin. Specifically, the binning step size for event energy statistics is defined as... , No. Time-divided into boxes for:

[0086] ;

[0087] in, `j` represents the start time of the common time interval, and `j` is the time bin identifier. (Statistics on the first...) Each observation perspective in the time bin The number of events within, denoted as This time is divided into boxes. The total energy E(j) within is defined as:

[0088] ;

[0089] In the formula, V represents the total number of observation viewpoints. The total energy sequence is obtained by arranging the total energy E(j) corresponding to all time bins in chronological order. That is, each time bin corresponds to one element in the total energy sequence. The total event count in each time bin under all observation viewpoints is used to represent the event energy in that time bin.

[0090] Step 2-2: Select sampling anchor points on the total energy sequence, take the time of the sampling anchor point as the starting time, and construct event fragment samples by extracting event segments from the event stream according to the window length of the fixed revolution window.

[0091] Selecting sampling anchor points: In one option, sampling anchor points are selected uniformly across the total energy sequence; in another option, sampling anchor points are selected from the high-energy candidate bins of the total energy sequence, wherein a high-energy threshold is first determined based on the total energy sequence. For example, the 95th percentile of the total energy sequence can be used as a candidate threshold, and combined with the minimum event count threshold for further limitation, i.e.: ,in, This represents the 95th percentile of the total energy sequence. This indicates the preset minimum event count threshold. This will satisfy... The time bins are used as high-energy candidate bins, and candidate sampling anchors are selected according to the event energy from high to low.

[0092] Furthermore, during the sample construction phase of model training, for multi-view original event streams labeled as healthy states, since they usually do not have obvious periodic impact responses, sampling anchors can be uniformly selected on the total energy sequence corresponding to the event streams; for multi-view original event streams labeled as fault states, sampling anchors can be selected from the high-energy candidate time bins corresponding to the event streams to increase the probability of including fault impact response event segments.

[0093] In some embodiments, the process of setting sampling anchor points preferably also includes applying dead zone constraints to adjacent sampling anchor points. These dead zone constraints refer to the requirement that two adjacent sampling anchor points maintain a time distance or bin spacing of not less than a preset minimum interval. Preferably, the minimum interval can be set as a preset ratio of the window duration of a fixed revolutions window to the number of bins, for example:

[0094] ;

[0095] in, This represents the dead zone ratio, with a preferred value of 0.5, though practical values ​​are empirical. When the distance between a new candidate sampling anchor point and an already selected sampling anchor point is less than the minimum interval... When this happens, the candidate sampling anchor point is removed or the one with higher energy is retained to reduce the high degree of overlap between adjacent samples.

[0096] On the other hand, the fixed rotational speed window is calculated; for the rotational speed (rpm) of the bearing under test, the single-rotation cycle and the window duration of the fixed rotational speed window can be expressed as:

[0097] ;

[0098] in, This indicates the rotational speed of the bearing under test, in revolutions per minute (rpm). Indicates a single revolution cycle, in seconds; This indicates the preset mechanical rotation speed, with a preferred value of 3. This represents the time length from which an event segment is extracted from the event stream, i.e., the window duration of the fixed-rotation window. It should be understood that the fixed-rotation window designed according to the above model ensures, on the one hand, a fixed predicted mechanical rotational speed, and on the other hand, is correlated with the actual rotational speed of the bearing under test.

[0099] In some embodiments, the process of constructing event fragment samples can also incorporate jitter offset, specifically:

[0100] A jitter offset is introduced near each sampling anchor point to generate multiple event fragment samples, improving sample diversity and temporal localization robustness. Specifically, for a sampling anchor point... The fixed rotation window duration is The basic window is: [ , + Set the jitter ratio. and in Generate one or more time offsets within the range Each offset window is: [ + , + + ], in ensuring the window Without exceeding a common time interval, corresponding event segments are extracted to form multiple event segment samples. Preferably, A value of 0.1 can be used, meaning the jitter range is 10% of the window duration. This jitter offset is used to increase the robustness of sample time localization without changing the number of mechanical revolutions covered by each sample. Therefore, the same anchor point can be used to obtain multiple event segments of the same length, all covering three revolutions, but with slightly different time positions, through different starting time offsets.

[0101] In the current implementation, the fixed revolution window is three revolutions. The duration of the three-revolution window is 180ms at 1000r / min and 90ms at 2000r / min. The window duration is different at different revolution rates, but the number of mechanical revolutions covered is the same, thus providing a basis for subsequent corner domain alignment.

[0102] The actual processing flow in this step is as follows: First, for the multi-view original event stream, construct the total energy sequence using the multi-view event counts within the common time interval; second, determine the sampling anchor point on the total energy sequence according to a uniform selection strategy or a high-energy interval selection strategy; third, based on the rotation speed of the sample to be measured... and preset speed Calculate the window duration for a fixed revolutions window Finally, using the sampling anchor point as the window positioning reference, the time length is extracted from the event streams of each observation perspective after decoding and synchronization in step 1. The event fragments are used as event fragment samples (abbreviated as event fragment samples). Subsequent differential enhancement event phase feature construction uses these event fragment samples as the processing object.

[0103] Step 3: Constructing the dual-channel event phase sequence.

[0104] Step 3-1: Divide the event segment samples from each observation perspective obtained in Step 2 into time slices. Let the preset time resolution for phase feature construction be... The preferred value is 50 microseconds. For the first... Event fragment samples from various observation perspectives ,according to It is divided into multiple consecutive time slices: ;

[0105] in, This represents the start time of a fixed-revolution window for the current event segment sample. This is the time slice number.

[0106] Step 3-2, for each event fragment sample Events within the same time slice are accumulated into a two-dimensional array according to their pixel coordinates to obtain the event frame corresponding to that time slice:

[0107] ;

[0108] in, Indicates the first The observation perspective in the first Pixels in an event frame formed by a time slice The cumulative event value at the location, This is a two-dimensional index variable representing the pixel position of the event frame currently being calculated. Indicates the first A sample of an event fragment from a single observation perspective; Indicates the first The event polarity of an event; This is an indicator function that takes the value 1 if the condition within the parentheses is true, and 0 otherwise. Indicates the first A time slice.

[0109] For a resolution of Event camera:

[0110] ;

[0111] calculate At that time, judge each one one by one. Within each time slice, the event coordinates of which events equal the current pixel position (x, y) are determined, and the polarities of these events are summed. Therefore, The range of values ​​is determined by the pixel resolution of the event camera and is a known discrete coordinate index.

[0112] For the i-th event in the v-th observation view, the indicator function is valid only if the following conditions are met simultaneously. Choose 1:

[0113] (1) This event belongs to the current event fragment sample. (2) The timestamp of the event Located in the nth time slice (3) The x-axis of the event (3) The x-coordinate of the current pixel; (4) The y-coordinate of the event. It equals the y-coordinate of the current pixel.

[0114] The meaning is: to determine the polarity of all events falling at pixel position (x, y) within the nth time slice. Accumulation is performed, with positive polarity events contributing +1 and negative polarity events contributing -1. That is, when the original polarity bit is 1, the mapping is... This indicates a brightness increase event; when the original polarity bit is 0, it is mapped to... This indicates an event where the brightness decreases.

[0115] Step 3-3: Construct the total event intensity and the differential response intensity in at least two directions. For each event fragment sample, the following feature construction is performed, and the final input into the model yields the corresponding diagnostic result. Therefore, the following sections will use event fragment samples as the data processing object. For example... Figure 3 The process shown in this embodiment employs differential response intensities in at least two directions, namely, transverse differential intensity and longitudinal differential intensity, whose expressions are as follows:

[0116] ;

[0117] ;

[0118] ;

[0119] in, and These represent the pixel width and height of the event camera, respectively; Indicates the first The observation perspective in the first The total event intensity of a time slice is used to reflect the overall level of event activity within that time slice. It represents the lateral differential intensity, used to reflect the local structural changes of the event frame along the horizontal direction; It represents the longitudinal differential intensity, used to reflect the local structural changes of the event frame along the vertical direction.

[0120] To reduce scale instability caused by event density fluctuations, this embodiment introduces an adaptive scaling factor based on median absolute deviation:

[0121] ;

[0122] ;

[0123] ;

[0124] in, Indicates the absolute deviation of the median. and All are minimal constants to prevent the denominator from being zero, and z is a user-defined parameter symbol; Indicates the first The combined differential response, consisting of transverse and longitudinal differential intensities, corresponding to each time slice. Indicates the first The integrated differential response sequence corresponding to all time slices from each observation perspective; For the first As shown in the formula above, each observation view has an independent adaptive scaling factor corresponding to the current fixed revolutions window. That is, to obtain a unified [response rate] for the "current fixed revolution window from the current observation perspective". First, calculate the time slices for each time slice within the fixed revolutions window. Then, the sequence composed of all time slices. Calculate ; This indicates that the value is restricted to a certain range. A truncation operator within the specified range. Preferably, , .

[0125] Steps 3-4: Based on the adaptive scaling factor, further construct the dual-channel event phase sequence:

[0126] ;

[0127] ;

[0128] ;

[0129] ;

[0130] in, and They represent the first The observation perspective in the first The horizontal and vertical event phases of each time slice; and All of these are minimal constants used to prevent the denominator from being zero; Indicates the first The number of time slices obtained by dividing the current fixed revolutions window from each observation perspective; and Indicates the first The observation perspective in the first The mean-reduced dual-channel event phases corresponding to each time slice are the mean-reduced horizontal event phase sequence and the vertical event phase. The above model is based on the following idea: constructing a dual-channel event phase path using the directional difference response after window-level MAD robust scaling as the first component and the total event activity intensity as the reference component.

[0131] Preferably, the dual-channel event phase sequence is further subjected to mean removal processing to obtain phase characterization results with stable numerical range and less affected by local noise.

[0132] It should be noted that although this embodiment uses lateral and longitudinal differential intensities to construct a dual-channel event phase sequence, in other embodiments, differential response intensities in other directions can also be used to construct the corresponding channels. For example, 45° directional differential response, 135° directional differential response, directional gradient response based on the Sobel operator, or radial and tangential differential responses constructed according to the geometric relationship of the bearing region under test can be used. As long as it can reflect the directional change characteristics of the event frame, it should fall within the protection scope of this invention, and the corresponding parameters of the lateral and longitudinal differential intensities mentioned above can be replaced.

[0133] Step 4: Corner domain resampling and multi-view adaptive fusion.

[0134] In this embodiment, since the duration of the fixed revolution window varies at different rotational speeds, the length of the dual-channel event phase sequence corresponding to each observation viewpoint may differ. Therefore, the dual-channel event phase sequences corresponding to each observation viewpoint are resampled to a uniform angular domain length to establish the angular domain correspondence between samples from different rotational speeds. The number of sampling points per revolution angular domain after resampling is... The preset rotation speed is Then the length of the unified angle domain is: .

[0135] Preferably, , ,therefore For the first From one observation perspective, the mean-removed horizontal event phase and vertical event phase sequence , Mapped to length of respectively through linear interpolation The angular domain coordinates are used to obtain the dual-channel angular domain phase sequence. , , where the corner domain identifier This process unifies time series of different lengths at different rotational speeds into a mechanical angular domain sequence of the same length.

[0136] In the application example, each revolution corresponds to 360 angular domain sampling points, and the total angular domain length for three revolutions is 1080 points. After resampling, dual-channel angular domain phase sequences for each observation view are obtained. Furthermore, to improve the stability of multi-view fusion, this embodiment determines the view fusion weights based on the impact significance of the dual-channel angular domain phase sequences corresponding to each observation view. Specifically, an intermediate representation is first constructed:

[0137] ;

[0138] Then calculate its Pearson kurtosis:

[0139] ;

[0140] ;

[0141] in, ;

[0142] In the formula, Indicates the first The standard deviation of the intermediate representation from each observation perspective This represents an approximately constant threshold value, with the preferred value being [value to be filled in]. .when At that time, it is assumed that the intermediate phase representation of the angular domain corresponding to the observation viewpoint is approximately constant, and its Pearson kurtosis is set to 0 to avoid instability in kurtosis calculation due to variance approaching 0. The specific setting process is as follows: Figure 4 The process is shown below.

[0143] The Pearson kurtosis of each observation viewpoint is transformed into the initial viewpoint fusion weights using Softmax:

[0144] ;

[0145] ;

[0146] in, and They represent the first The observation perspective in the first Two angular domain phase channels at each angular domain sampling point; For the first Intermediate representation of current event fragment samples from various observation perspectives; The mean of the intermediate representations; , All are Pearson kurtosis, used to characterize the impact significance of angular domain phase sequences; The weight distribution control parameter is preferably set to 2.0; Weigh the initial perspectives; The smoothing coefficient is preferably set to 0.05. Weights are integrated for the final perspective; The number of observation viewpoints; when the intermediate representation corresponding to a certain observation viewpoint is approximately constant, the kurtosis of that observation viewpoint is set to 0 to avoid weak response viewpoints from receiving excessive weight.

[0147] It should be noted that the dual-channel corner domain input designed based on the above data model has the following characteristics:

[0148] (1) Sample-level dynamism: The weights in the above model are not fixed after training, nor are they set to a uniform ratio for the entire dataset. Instead, they are calculated separately for each event segment sample.

[0149] (2) Impact sensitivity: This invention fully considers that local bearing failures often form sparse and prominent impact peaks in the event phase sequence. Kurtosis can reflect the significance of this non-Gaussian impact. Therefore, it is proposed that: first, the two angular domain phase channels of the same observation angle are averaged to form an intermediate characterization, and then the Pearson kurtosis is calculated.

[0150] (3) Viewpoint reliability mapping: Softmax with control parameter β is used to convert the kurtosis of different viewpoints into normalized weights, so that the viewpoints with obvious shock response can make a higher contribution.

[0151] (4) Preventing single-view monopoly: Add to the Softmax weights A smooth lower bound is used to avoid giving too much weight to one perspective while retaining complementary information from other perspectives.

[0152] (5) Abnormal viewpoint suppression: When the intermediate representation of a certain viewpoint is approximately constant, its kurtosis is set to 0, which can prevent the viewpoint with no effective response or near dead line from gaining high weight due to numerical abnormalities.

[0153] Finally, the dual-channel angular domain phase sequences corresponding to each observation viewpoint are weighted and summed using the final viewpoint fusion weights to obtain the fused dual-channel angular domain input X:

[0154] ;

[0155] ;

[0156] In the formula, , These are the two channel variables of the fused dual-channel angular domain input X, with dimensions of... .

[0157] Step 5: Fault diagnosis output.

[0158] In this embodiment, the dual-channel corner domain of each fused event fragment sample is input. The fault diagnosis model is input to output the fault category of the bearing under test. The fault diagnosis model includes a angular domain temporal backbone branch, a statistical prior generation branch, a conditional gating module, and a classification head. Dual-channel angular domain input. Input the angular domain temporal trunk branch, and statistical prior generation branches are generated from the dual-channel angular domain input of the current event segment sample. Extract kurtosis, peak value, and skewness to form a six-dimensional event-angle domain statistical prior vector. .

[0159] The statistical prior generation branches are generated from the dual-channel corner domain inputs of the current event fragment sample. Extract kurtosis, peak value, and skewness to form a six-dimensional event-angle domain statistical prior vector. :

[0160] ;

[0161] in, and Used to characterize the impact spike intensity of the two angular domain phase channels. and Used to characterize the maximum transient response of the two angular domain phase channels. and Used to characterize the asymmetry of the distribution of phase channels in two angular domains, T is the matrix transpose symbol.

[0162] The statistical prior is derived from the dual-channel angular domain input of the event segment sample currently being diagnosed, specifically:

[0163] 1. First, obtain the fused dual-channel corner domain input of the current sample;

[0164] 2. Calculate the kurtosis, peak value, and skewness of the two channels for the current sample respectively;

[0165] 3. Form a six-dimensional event-domain statistical prior vector for the current sample;

[0166] 4. Use this six-dimensional event field to statistically analyze the prior vector. Generate the conditional gating vector corresponding to the current event segment sample;

[0167] 5. Network features used to modulate current event fragment samples.

[0168] "Prior" refers to its physical prior meaning in terms of bearing impact mechanism and statistical shape, not to something derived from a historical database.

[0169] In this embodiment, the corner-domain temporal backbone branch includes three sequentially connected residual Mamba modules; for the output features of each residual Mamba module, the conditional gating module utilizes the six-dimensional event corner-domain statistical prior vector. A corresponding 64-dimensional conditional gating vector is generated, and the state-space feature output of the residual Mamba module is modulated channel by channel using the conditional gating vector. Specifically:

[0170] For the The output features of the residual Mamba module, the conditional gating module first statistically analyzes the prior vectors in the event angle domain. Transformed into a conditional gating vector through linear mapping and the Sigmoid function. :

[0171] ;

[0172] in, , This represents the number of hidden feature channels in the residual Mamba module; and They represent the first Trainable weights and biases for a conditional gating mapping.

[0173] The first residual Mamba module takes a dual-channel angular domain input. or The features after projection through the linear mapping layer are defined as follows: The input to each residual Mamba module is Its state-space feature output is:

[0174] ;

[0175] In the formula, Let represent the state-space feature transformation function corresponding to the i-th Mamba module. It receives the input angular domain sequence features of the i-th module, and through the selective state-space sequence modeling process of Mamba, outputs the state-space features, where j is the residual Mamba module number. The conditional gating vector... Broadcasting along the angular domain sequence (a common dimensional expansion operation in neural network tensor computation) yields the gated matrix. And perform conditional gating residual fusion according to the following formula:

[0176] ;

[0177] in, This represents element-wise multiplication. Therefore, the introduced event-angle domain statistical prior can adjust the state-space feature transfer intensity of each latent feature channel of the residual Mamba module based on the impact spike intensity, transient response strength, and phase distribution asymmetry of the current sample.

[0178] This invention uses event phase representation, corner domain alignment, and multi-view adaptive fusion as the foundation for front-end feature construction. Furthermore, it utilizes the statistical prior formed by the same dual-channel corner domain input to conditionally modulate the state-space feature transfer process within the residual Mamba module. Therefore, the front-end feature processing result (dual-channel corner domain input) is... This not only serves as input to the fault diagnosis model but also controls the feature extraction process within the model, enabling continuous data dependencies between event phase construction, angular domain statistical priors, and long-sequence state space modeling. In the current preferred embodiment, the fused dual-channel angular domain input... First, the feature is projected onto a 64-dimensional latent feature space through a linear mapping layer to obtain the initial angular domain sequence features, which are then input into the first residual Mamba module.

[0179] It should be understood that the same It has two utilization paths: on the one hand, After feature encoding, the input is given to the residual Mamba backbone for extracting long sequence features in the angular domain; on the other hand, statistical priors generate branches from the current sample. The kurtosis, peak value, and skewness of the two channels are calculated to form an event angular domain statistical prior vector. This prior vector is then used to generate a conditional gating vector for feature transfer in the modulated residual Mamba module. Therefore, it is not a dual-channel angular domain input. It does not directly control the network, but rather its derived statistical priors participate in the network's internal modulation.

[0180] In this embodiment, the first residual Mamba module extracts the corner domain impulse response and long-range phase dependence at a full 1080-point corner domain resolution. After conditionally gated residual fusion, max pooling is used to compress the sequence length from 1080 to 540 to reduce the computational cost of subsequent state space modeling. The second and third residual Mamba modules continue to perform conditionally gated state space feature extraction on the compressed corner domain sequence.

[0181] After processing by the third residual Mamba module, global average pooling is performed on the time dimension to obtain 64-dimensional sample-level angular domain time-series features. Further, the 64-dimensional angular domain time-series features are concatenated with the six-dimensional event angular domain statistical prior vector and input into a two-layer fully connected classification head. The probability of the bearing under test belonging to a healthy state, inner race fault, outer race fault, or rolling element fault is output through Softmax.

[0182] It should be understood that each event segment sample yields a fault category result. For the original event stream to be diagnosed, multiple event segment samples will also be obtained. Finally, the multiple segments of the same event stream can be averaged to obtain the final fault category of the event stream. The network can then be trained and network parameters adjusted using labels and loss functions. This part can refer to existing technologies, and this invention does not impose specific limitations on it.

[0183] In practical applications, the trained network model is used to average the class probabilities of multiple event segment samples from the same event stream to obtain the final fault category of the event stream. Alternatively, event segments can be arranged in chronological order according to the original event stream to form a continuous segment diagnosis sequence. If different diagnosis results exist, a weighted average can be used.

[0184] Example 2:

[0185] This embodiment provides a diagnostic system based on the above-described fault diagnosis method, comprising at least:

[0186] The multi-view event stream acquisition and synchronization module is used to acquire event stream data of the bearing under test from at least two observation views through the event camera and to perform common time interval clipping on the event streams from different observation views.

[0187] The sample construction module is used to extract event fragments from the event stream as event fragment samples according to the window length of a fixed revolutions window. The fixed revolutions window is designed based on the bearing speed and meets the requirement of covering the mechanical revolutions. Fixed dynamic duration window;

[0188] The dual-channel event phase sequence construction module is used to divide each event segment sample into time slices, and then construct event frames by accumulating events in the same time slice. The total event intensity and differential response intensity in at least two directions are then calculated, and the dual-channel event phase sequence is constructed using the total event intensity and differential response intensity.

[0189] The angular domain resampling and multi-view adaptive fusion module is used to resample the dual-channel event phase sequence corresponding to each observation view to a unified angular domain length to obtain a dual-channel angular domain phase sequence; then, the dual-channel angular domain phase sequences of all observation views are fused to obtain the fused dual-channel angular domain input.

[0190] The diagnostic output module is used to send the dual-channel angular domain input to the fault diagnosis model and output the fault category of the bearing.

[0191] For the specific implementation process of each module, please refer to the above method content. This invention will not repeat it here. The above division of functional modules is only for illustrative purposes. In some embodiments, some functional modules can be merged and some functional modules can be split. Each functional module can be implemented in software, hardware, or a combination of software and hardware. Among them, software and hardware devices include, but are not limited to, general-purpose computer equipment, programmable gate arrays, digital signal processors, microprocessors and their corresponding programming or burning software.

[0192] Example 3:

[0193] This embodiment provides a computer terminal, comprising at least one or more processors and a memory storing one or more computer programs. The processors invoke the computer programs to load a pre-trained fault diagnosis model or construct and train a fault diagnosis model, acquire event signals of the bearing to be monitored and perform decoding, fixed-rotation event segment extraction, event phase feature construction, angular domain resampling, and multi-view adaptive fusion. An event angular domain statistical prior vector is generated from the fused dual-channel angular domain input. A conditional gating vector is generated based on the statistical prior vector. The conditional gating vector is used to modulate the state-space feature output of the residual Mamba module, and fault diagnosis is completed based on the modulated angular domain long sequence features. It should be understood that in this embodiment, the processor may be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. Memory can include read-only memory and random access memory, providing instructions and data to the processor. A portion of the memory can also include non-volatile random access memory. For example, memory can also store device type information.

[0194] Example 4:

[0195] This embodiment provides a computer-readable storage medium storing a computer program that is called by a processor to perform the following: loading a pre-trained fault diagnosis model or constructing and training a fault diagnosis model; acquiring event signals of the bearing to be monitored, decoding them, and extracting event phase features; performing angular domain resampling and multi-view adaptive fusion; generating event angular domain statistical prior vectors and corresponding conditional gating vectors; using the conditional gating vectors to modulate the state space features output by the residual Mamba module; and completing fault diagnosis based on the modulated angular domain long sequence features.

[0196] Please refer to the explanation of the method above for the specific implementation process of each step.

[0197] The readable storage medium is a computer-readable storage medium, which can be an internal storage unit of the hardware and software device described in any of the foregoing embodiments, such as the hard drive or memory of the controller. The readable storage medium can also be an external storage device of the controller, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, etc., equipped on the controller. Further, the readable storage medium can include both internal storage units and external storage devices of the controller. The readable storage medium is used to store the computer program and other programs and data required by the controller. The readable storage medium can also be used to temporarily store data that has been output or will be output.

[0198] Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned readable storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0199] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application refers to flowchart illustrations and / or instructions executed by a processor of a method, apparatus (system), and computer program product according to embodiments of this application to create means for implementing the functions specified in one or more flowchart illustrations and / or one or more block diagrams. These computer program instructions may also be stored in a computer-readable storage medium capable of directing a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowchart illustrations and / or one or more block diagrams. These computer program instructions may 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 functions specified in one or more flowcharts and / or one or more blocks of a block diagram.

[0200] It should be emphasized that the examples described in this invention are illustrative rather than limiting. Therefore, this invention is not limited to the examples described in the specific embodiments. Any other embodiments derived by those skilled in the art based on the technical solutions of this invention, without departing from the spirit and scope of this invention, whether modifications or substitutions, are also within the protection scope of this invention.

Claims

1. A bearing cross-speed fault diagnosis method based on angular domain phase representation driven by an event camera, characterized in that, Includes the following steps: Step 1: Multi-view event stream acquisition and synchronization, that is, acquiring event stream data of the bearing under test from at least two observation perspectives through the event camera and clipping the event streams from different observation perspectives to a common time interval. Step 2: Sample Construction. Event segments are extracted from the event stream according to the window length of a fixed revolutions window as event segment samples. The fixed revolutions window is designed based on the bearing speed and meets the requirement of covering the mechanical revolutions. Fixed dynamic duration window; Step 3: Constructing a dual-channel event phase sequence. For each event segment sample, time slices are divided, and event frames are constructed by accumulating events in the same time slice. The total event intensity and differential response intensity in at least two directions are then calculated. The dual-channel event phase sequence is constructed using the total event intensity and differential response intensity. Step 4: Corner domain resampling and multi-view adaptive fusion. The dual-channel event phase sequence corresponding to each event fragment sample under each observation view is resampled to a uniform corner domain length to obtain a dual-channel corner domain phase sequence; then the dual-channel corner domain phase sequences of all observation views are fused to obtain the dual-channel corner domain input after event fragment sample fusion. Step 5: Send the dual-channel angular domain input of the event fragment sample to the fault diagnosis model to obtain the fault classification probability of the event fragment sample, and then combine the fault classification probabilities of the event fragment samples on the same event stream to obtain the bearing fault category.

2. The method according to claim 1, characterized in that: Step 3, which involves constructing a dual-channel event phase sequence using the total event intensity and the differential response intensity, can be represented by the following mathematical model: ； ； ； ； In the formula, Indicates the first The observation perspective in the first The total event intensity of each time slice is used to reflect the overall level of event activity within the time slice; Indicates the first The observation perspective in the first A type of differential intensity for each time slice; Indicates the first The observation perspective in the first Another type of differential intensity for a time slice; and They represent the first The observation perspective in the first One type of event phase and another type of event phase in a time slice; and All of these are minimal constants used to prevent the denominator from being zero; Indicates the first The number of time slices obtained by dividing the current fixed revolutions window from each observation perspective; and This represents the mean-free dual-channel event phase sequence, corresponding to one type of event phase and another type of event phase; For the first The adaptive scaling factor corresponding to the current fixed revolutions window for each observation view; It is the arctangent function in the four quadrants.

3. The method according to claim 1, characterized in that: Adaptive scaling factor An adaptive scaling factor based on median absolute deviation is used, and each fixed-rotation window under each observation view corresponds to one adaptive scaling factor. Specifically: ； ； ； in, Indicates the absolute deviation of the median. and All are minimal constants to prevent the denominator from being zero, and z is a user-defined parameter symbol; Indicates the first The observation perspective in the first The combined differential response consisting of two types of differential intensities in each time slice; Indicates the first The composite differential response sequence consisting of all time slices from each observation perspective; Indicates the first The total event intensity sequence consisting of all time slices from each observation perspective; This indicates that the value is restricted to a certain range. Truncation operator within range, and These are the minimum and maximum values ​​of the preset scaling factor, respectively.

4. The method according to claim 1, characterized in that: Step S3 involves dividing each event segment sample into time slices and then accumulating events within the same time slice to construct an event frame. An event frame is defined as follows: ； in, Indicates the first The observation perspective in the first Pixels in an event frame formed by a time slice The cumulative event value at the location; Indicates the first A sample of an event fragment from a single observation perspective; Indicates the first The event polarity of an event; This is an indicator function that takes the value 1 if the condition within the parentheses is true, and 0 otherwise. Indicates the first A time slice; For the i-th event in the v-th observation view, the indicator function is valid only if the following conditions are met simultaneously. Choose 1: (1) The event belongs to the current event fragment sample. ; (2) Event timestamp Located in the nth time slice Inside; (3) The x-axis of the event It equals the x-coordinate of the current pixel; (4) The vertical axis of the event It equals the y-coordinate of the current pixel.

5. The method according to claim 1, characterized in that: The formula for calculating the fixed revolutions window in step 2 is: ； in, This indicates the bearing speed, measured in revolutions per minute (rpm). Indicates a single revolution cycle, in seconds; This indicates the duration of the event segment extracted from the event stream, i.e., the window duration of a fixed-revolution window; The process of generating the event fragment samples is as follows: within the common time interval, the event counts of each observation perspective are counted with a bin step size to obtain the event energy in each time bin, thereby forming a total energy sequence; sampling anchor points are selected on the total energy sequence, and the time of the sampling anchor point is taken as the starting time. Event fragments are extracted from the event stream according to the window length of a fixed number of revolutions window to construct the event fragment samples.

6. The method according to claim 1, characterized in that: In step 4, the viewpoint fusion weight is determined based on the impact significance of the dual-channel angular domain phase sequences corresponding to each observation viewpoint. Then, the dual-channel angular domain phase sequences of all observation viewpoints are fused using the viewpoint fusion weight. Specifically: First, construct intermediate representations: ； Then, Pearson kurtosis is calculated, and the Pearson kurtosis of each observation view is transformed into initial view fusion weights using Softmax, and then converted into final view fusion weights: ； ； in, and They represent the first The observation perspective in the first The two angular domain phase channels at each angular domain sampling point are obtained by resampling the dual-channel event phase sequence corresponding to each observation view to a uniform angular domain length. For the first Intermediate representation of current event fragment samples from various observation perspectives; , , respectively, are the Pearson kurtosis corresponding to the v-th and r-th observation viewpoints, used to characterize the impact significance in the angular domain phase sequence; These are the control parameters for the weight distribution; Weigh the initial perspectives; For smoothing coefficients; Weights are integrated for the final perspective; The fusion model is as follows: ； ； in, , These are the two channel variables of the fused dual-channel angular domain input X, with dimensions of... ; This represents the number of observation angles.

7. The method according to claim 1, characterized in that: The fault diagnosis model includes a angular domain temporal backbone branch, a statistical prior generation branch, a conditional gating module, and a classification head; the dual-channel angular domain input is... The input angular domain temporal backbone branch and the statistical prior generation branch are respectively used. The statistical prior generation branch utilizes the dual-channel angular domain input of the current event segment sample. Generate event-angle domain statistical prior vectors The conditional gating module will statistically analyze the prior vectors of the event angle domain. Mapped to a conditional gating vector; The angular domain time-series backbone branch has at least a plurality of sequentially connected residual Mamba modules. For each residual Mamba module, the angular domain long sequence features output by the residual Mamba module are modulated channel by channel using the conditional gating vector. The modulated features corresponding to the last residual Mamba module are processed by the classification head and output as the fault classification probability of each event segment sample. The fault classification result is then obtained by combining the fault classification probabilities of event segment samples on the same event stream.

8. A diagnostic system based on the method of any one of claims 1-7, characterized in that: Includes the following modules: The multi-view event stream acquisition and synchronization module is used to acquire event stream data of the bearing under test from at least two observation views through the event camera and to perform common time interval clipping on the event streams from different observation views. The sample construction module is used to extract event fragments from the event stream as event fragment samples according to the window length of a fixed revolutions window. The fixed revolutions window is designed based on the bearing speed and meets the requirement of covering the mechanical revolutions. Fixed dynamic duration window; The dual-channel event phase sequence construction module is used to divide each event segment sample into time slices, and then construct event frames by accumulating events in the same time slice. The total event intensity and differential response intensity in at least two directions are then calculated, and the dual-channel event phase sequence is constructed using the total event intensity and differential response intensity. The corner domain resampling and multi-view adaptive fusion module is used to resample the dual-channel event phase sequence corresponding to each event segment sample under each observation view to a unified corner domain length to obtain a dual-channel corner domain phase sequence; then, the dual-channel corner domain phase sequences of all observation views are fused to obtain the dual-channel corner domain input after the event segment sample fusion. The diagnostic output module is used to send the dual-channel angular domain input of the event fragment samples to the fault diagnosis model to obtain the fault classification probability of the event fragment samples, and then combine the fault classification probabilities of the event fragment samples on the same event stream to obtain the bearing fault category.

9. A computer device, characterized in that: include: One or more processors; A memory that stores one or more computer programs; The processor calls the computer program to implement: The steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1-7.

Images