A fault diagnosis method, apparatus, equipment and medium for a triaxial vibration sensor

Abstract

This invention relates to the field of fault diagnosis technology, and discloses a fault diagnosis method, apparatus, device, and medium for a triaxial vibration sensor. The method includes: acquiring vibration signals collected by the triaxial vibration sensor during the operation of a target device; extracting the temporal features of each axis's vibration signal, and constructing a triaxial relationship matrix by using the dot product similarity or the similarity after linear mapping between any two axis temporal features to characterize the statistical correlation and structural consistency between the triaxial vibration signals; using each axis's temporal feature as a graph node, performing graph convolution aggregation on the time features of each axis according to the triaxial relationship matrix, and weighting and aggregating the structural reference information of the other two axes to the current axis to obtain the graph features of each axis's vibration signal; fusing the graph features of each axis with the temporal features to obtain a fused feature containing the individual features of the corresponding axis's vibration signal and the common features of the three axes; and performing fault diagnosis on the triaxial vibration sensor using the fused feature of the triaxial vibration signal.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and in particular to a fault diagnosis method, apparatus, equipment and medium for a triaxial vibration sensor. Background Technology

[0002] Triaxial vibration sensors are widely used in condition monitoring scenarios such as rotating machinery, pump units, bearings, and traction drive systems. Compared to single-axis sensors, triaxial vibration sensors can simultaneously acquire vibration responses in three directions, which helps improve the completeness and fault sensitivity of vibration monitoring. However, during long-term operation, the sensors themselves may experience faults such as fixed deviations, drift, gain distortion, jamming, and noise pollution. When the sensor output is abnormal, downstream equipment condition monitoring systems may easily misinterpret the sensor abnormality as a fault in the equipment itself, leading to false alarms, unnecessary shutdowns, or wasted maintenance resources. Therefore, fault diagnosis of triaxial vibration sensors is of great significance for system condition monitoring.

[0003] Existing fault diagnosis methods for triaxial vibration sensors include fault diagnosis schemes based on simple single-axis or multi-axis splicing and anomaly identification schemes based on fixed rules or empirical thresholds. Fault diagnosis schemes based on simple single-axis or multi-axis splicing first acquire vibration signals, then extract features such as mean, variance, kurtosis, and spectral energy, or directly input the raw time window into models such as convolutional neural networks or recurrent neural networks, with the classifier outputting categories such as normal, bias, drift, gain, stuck, or noise. Anomaly identification schemes based on fixed rules or empirical thresholds set thresholds based on the amplitude, mean drift, signal energy, or noise intensity of a certain axis; when the threshold is exceeded, the sensor is considered abnormal.

[0004] The common feature of the above-mentioned existing solutions is that most of them only utilize the local signal changes of each axis itself. Even when using triaxial data, they usually just cascade the signals in the three directions into the network. They do not make full use of the relationship between the triaxial vibration responses, resulting in insufficient utilization of the structural abnormal changes caused by single-axis faults, thus affecting the accuracy of fault diagnosis. Summary of the Invention

[0005] The purpose of this invention is to provide a fault diagnosis method, device, equipment, and medium for triaxial vibration sensors, which can solve the problem of low fault diagnosis accuracy.

[0006] To address the aforementioned technical problems, embodiments of the present invention provide a fault diagnosis method for a triaxial vibration sensor, comprising the following steps: Acquire vibration signals from the three axes of the triaxial vibration sensor during the operation of the target equipment; The temporal features of each axis vibration signal are extracted, and the relationship score between any two axis vibration signals is determined by the dot product similarity or the similarity after linear mapping between the temporal features of any two axes, so as to construct a triaxial relationship matrix. The triaxial relationship matrix is ​​used to characterize the statistical correlation and structural consistency between the triaxial vibration signals of the triaxial vibration sensor. Using the temporal features of each axis vibration signal as graph nodes, graph convolution aggregation is performed on the temporal features of each axis according to the three-axis relationship matrix to weight and aggregate the structural reference information of the other two axes to the current axis, so as to obtain the graph features of each axis vibration signal. The graph features of each axis vibration signal are then fused with the temporal features to obtain the fused features that include the individual features of the corresponding axis vibration signal and the common features of the three axes. Fault diagnosis of triaxial vibration sensors is performed by fusing the characteristics of triaxial vibration signals.

[0007] Furthermore, the fault diagnosis results of the triaxial vibration sensor are obtained by inputting the vibration signals collected from the three axes respectively into a trained fault diagnosis model; wherein, the fault diagnosis model includes a temporal feature extraction layer, a graph attention learning layer, a graph feature fusion layer, and a main classification layer; The temporal feature extraction layer uses a three-parameter-shared one-dimensional convolutional coding network to map the vibration signal of each axis to the corresponding temporal features; the one-dimensional convolutional coding network contains three one-dimensional convolutional modules; The graph attention learning layer uses the temporal features of each axis vibration signal as graph nodes to construct a three-axis adaptive relationship graph. The relationship score between any two axis vibration signals is determined by the dot product similarity or the similarity after linear mapping between the temporal features of any two axes in the three-axis adaptive relationship graph. Then, the corresponding relationship weight is obtained by normalizing each relationship score to construct a three-axis relationship matrix. The graph feature fusion layer performs graph convolution aggregation on the temporal features of each axis vibration signal based on the triaxial relationship matrix, so as to weight and converge the structural reference information of the other two axes to the current axis to obtain the graph features of each axis vibration signal. The graph features of each axis vibration signal are then concatenated with the temporal features, weighted summation or gated fusion is performed to obtain the corresponding fused features. Finally, the fused features of the triaxial vibration signals are concatenated and fully connected to obtain the sample-level diagnostic representation. The main classification layer outputs the probability distribution of the triaxial vibration sensor for various different states based on the sample-level diagnostic characterization, and takes the state with the highest probability as the fault diagnosis result of the triaxial vibration sensor.

[0008] Furthermore, the fault diagnosis of the triaxial vibration sensor includes: If, based on the probability distribution of the triaxial vibration sensor with respect to various different states, it is determined that there are multiple probabilities greater than a preset threshold, then a relationship heatmap is generated based on the change in the relationship strength of each axis in the triaxial relationship matrix. The abnormal axis in the triaxial vibration sensor is determined based on the relationship heatmap. If the relationship between one axis and the other two axes of the triaxial vibration sensor weakens, it indicates that the relationship between the current axis and the other two axes deviates from the normal triaxial relationship pattern and is determined to be an abnormal axis.

[0009] Furthermore, the fault diagnosis model is trained in the following manner: At least one axis vibration signal is randomly selected for time feature or fusion feature masking, and the masked features are recovered through a pre-constructed auxiliary reconstruction network. The auxiliary classification head is used to assist in the fault diagnosis of the triaxial vibration sensor based on the recovered features. The main classification loss function is constructed based on the error between the output of the main classification layer and the true state of the triaxial vibration sensor; the mask reconstruction loss function is constructed based on the error between the features recovered by the auxiliary reconstruction network and the true features before masking; and the auxiliary classification loss function is constructed based on the error between the output of the auxiliary classification head and the true state of the triaxial vibration sensor. A fault diagnosis model is trained using the main classification loss function, the mask reconstruction loss function, and the auxiliary classification loss function.

[0010] Furthermore, the sample data used during the training of the fault diagnosis model includes samples in various different states: normal state samples, fixed deviation fault samples, drift fault samples, gain fault samples, jammed fault samples, and noise fault samples. The main classification layer outputs the probability distribution of the triaxial vibration sensor for various states such as normal state, fixed deviation fault, drift fault, gain fault, jamming fault and noise fault based on the sample-level diagnostic characterization, and takes the state with the highest probability as the fault diagnosis result of the triaxial vibration sensor.

[0011] Furthermore, the samples of the various different states are generated according to the following fault generation rules: Normal state samples are generated by collecting vibration signals from a triaxial vibration sensor in a healthy state, with the other two directions remaining unchanged; fixed deviation fault samples are generated by superimposing a constant bias onto the vibration signal of the target axis of the triaxial vibration sensor; drift fault samples are generated by superimposing a low-frequency trend term that changes over time onto the vibration signal of the target axis of the triaxial vibration sensor; gain fault samples are generated by scaling or amplifying the amplitude of the vibration signal of the target axis of the triaxial vibration sensor; jamming fault samples are generated by replacing the vibration signal of the target axis of the triaxial vibration sensor, either locally or globally, with constant or repetitive values; noise fault samples are generated by superimposing random noise and local burst noise onto the vibration signal of the target axis of the triaxial vibration sensor.

[0012] Furthermore, prior to extracting the temporal characteristics of the vibration signal per axis, the method further includes: The vibration signal of each axis is segmented into sliding windows according to a preset window length and a preset step size to obtain vibration signals of multiple different time windows for each axis. Then, the vibration signals of multiple different time windows for each axis are subjected to preprocessing operations of mean removal and standard deviation normalization.

[0013] Embodiments of the present invention also provide a fault diagnosis device for a triaxial vibration sensor, comprising the following modules: The signal acquisition module is used to acquire the vibration signals collected by the triaxial vibration sensor along the three axes during the operation of the target device. The relationship construction module is used to extract the temporal features of each axis vibration signal and determine the relationship score between any two axis vibration signals by the dot product similarity or the similarity after linear mapping between the temporal features of any two axes, so as to construct a three-axis relationship matrix. The three-axis relationship matrix is ​​used to characterize the statistical correlation and structural consistency between the three-axis vibration signals of the three-axis vibration sensor. The feature extraction module is used to take the time features of each axis vibration signal as graph nodes, perform graph convolution aggregation on the time features of each axis according to the three-axis relationship matrix, and weight and aggregate the structural reference information of the other two axes to the current axis to obtain the graph features of each axis vibration signal. The graph features of each axis vibration signal are then fused with the time features to obtain the fused features that include the individual features of the corresponding axis vibration signal and the common features of the three axes. The fault diagnosis module is used to diagnose faults in triaxial vibration sensors by using the fusion characteristics of triaxial vibration signals.

[0014] Embodiments of the present invention also provide a computer device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described fault diagnosis method for a triaxial vibration sensor.

[0015] Embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described fault diagnosis method for a triaxial vibration sensor.

[0016] The fault diagnosis method for triaxial vibration sensors provided by this invention has at least the following beneficial effects: By constructing a triaxial relationship matrix, the statistical correlation and structural consistency between the vibration signals collected by the triaxial vibration sensor on each of the three axes during the operation of a mechanical device are obtained. Then, by fusing the relationship graph features of the triaxial vibration signals, the individual characteristics of the triaxial vibration signals and the common characteristics of the triaxial vibration signals are obtained. The individual characteristics of the triaxial vibration signals represent the waveform characteristics of the corresponding axis vibration signals of the triaxial vibration sensor in the time domain, while the common characteristics of the triaxial vibration signals reflect the structural relationship characteristics between the triaxial vibration signals. Therefore, not only can the complete features that can reflect the current state (fault condition) of the triaxial vibration sensor be captured, but the abnormal changes caused by the single-axis fault of the triaxial vibration sensor to the overall relationship structure can also be effectively captured, thereby improving the fault diagnosis accuracy of the triaxial vibration sensor. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0018] Figure 1 A schematic flowchart of a fault diagnosis method for a triaxial vibration sensor provided by the present invention; Figure 2 A schematic diagram of the training process for a fault diagnosis model provided by the present invention; Figure 3 A schematic diagram of the framework structure of a fault diagnosis model provided by the present invention; Figure 4 This is an abnormal schematic diagram of a faulty axis relationship provided by the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0020] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0021] One embodiment of the present invention relates to a fault diagnosis method for a triaxial vibration sensor. The specific flow of the fault diagnosis method for the triaxial vibration sensor in this embodiment is as follows: Figure 1 As shown, it includes: Step 101: Obtain the vibration signals collected by the three axes of the triaxial vibration sensor during the operation of the target device.

[0022] Step 102: Extract the temporal features of each axis vibration signal, and determine the relationship score between any two axis vibration signals by the dot product similarity or the similarity after linear mapping between the temporal features of any two axes, so as to construct a triaxial relationship matrix; wherein, the triaxial relationship matrix is ​​used to characterize the statistical correlation and structural consistency between the triaxial vibration signals of the triaxial vibration sensor.

[0023] Step 103: Using the time features of each axis vibration signal as graph nodes, perform graph convolution aggregation on the time features of each axis according to the three-axis relationship matrix to weight and aggregate the structural reference information of the other two axes to the current axis, obtain the graph features of each axis vibration signal, and fuse the graph features of each axis vibration signal with the time features to obtain the fused features containing the individual features of the corresponding axis vibration signal and the common features of the three axes.

[0024] Step 104: Perform fault diagnosis on the triaxial vibration sensor by analyzing the fusion characteristics of the triaxial vibration signals. The following is a detailed description of the implementation details of the fault diagnosis method for the triaxial vibration sensor in this embodiment. The following content is only for the convenience of understanding and is not necessary for implementing this solution.

[0025] This embodiment outputs the fault diagnosis results of the triaxial vibration sensor by inputting the vibration signals acquired from the three axes into a trained fault diagnosis model. The model training process for the fault diagnosis model is as follows: Figure 2 As shown.

[0026] (1) Data collection and sample construction: First, triaxial vibration signals of the object under test during operation are acquired, including vibration responses in the X, Y, and Z axes. In practice, a triaxial vibration sensor installed at the same measuring point on the equipment is used to simultaneously acquire vibration sequences in the three directions.

[0027] Then, the continuously acquired triaxial vibration sequence is sliced ​​into sliding window segments according to a preset window length L, with a sliding step size of . Each time, a three-axis time window sample of length L is extracted from the original sequence to form a sample matrix. In one specific embodiment, the window length L can be set to 4096 points, with a step size of... The redundancy level of the sampling is preset.

[0028] Next, mean and standard deviation normalization is performed on each window sample to reduce the impact of differences in absolute dimensions between different samples and to preserve information on the time pattern and relative statistical structure of the three axes under the current operating conditions. In some embodiments, in addition to window normalization, wavelet preprocessing, bandpass filtering, or a combination thereof may also be used for processing.

[0029] In one example, this embodiment first performs file-level partitioning based on the original data file, and then performs the aforementioned sliding window sampling on each partition result to form a training set, a validation set, and a test set. Samples from different files are not cross-referenced into different sets to avoid sample leakage. The training set is used for parameter learning, the validation set is used to monitor model convergence and select optimal parameters, and the test set is used to evaluate the model's diagnostic performance.

[0030] In some embodiments, the sample data during fault diagnosis model training includes samples in several different states: normal state samples, fixed deviation fault samples, drift fault samples, gain fault samples, jammed fault samples, and noise fault samples.

[0031] Therefore, this embodiment will use healthy triaxial time window samples as a basis to generate samples of the above-mentioned various states according to the following fault generation rules: normal state samples are generated by vibration signals collected by triaxial vibration sensors in a healthy state; fixed deviation fault samples are generated by superimposing a constant bias on the vibration signal of the target axis of the triaxial vibration sensor; drift fault samples are generated by superimposing a low-frequency trend term that changes with time on the vibration signal of the target axis of the triaxial vibration sensor; gain fault samples are generated by scaling or amplifying the amplitude of the vibration signal of the target axis of the triaxial vibration sensor; jamming fault samples are generated by replacing the local or overall vibration signal of the target axis of the triaxial vibration sensor with constant or repetitive values; noise fault samples are generated by superimposing random noise and local burst noise on the vibration signal of the target axis of the triaxial vibration sensor.

[0032] (2) Fault diagnosis model construction: The fault diagnosis model framework structure in this embodiment is as follows: Figure 3 As shown, it includes a temporal feature extraction layer, a graph attention learning layer, a graph feature fusion layer, and a main classification layer.

[0033] The temporal feature extraction layer employs a three-parameter-shared one-dimensional convolutional coding network to map each axis vibration signal to its corresponding temporal feature. The three-axis vibration time windows are denoted as follows: , , This corresponds to the X-axis, Y-axis, and Z-axis signals. A one-dimensional convolutional coding network with shared parameters is used to extract temporal features in the three directions, resulting in... , and The shared weight design, based on the fact that the three axes belong to the same mechanical system and are observed from multiple directions, can reduce the number of model parameters and prevent overfitting while maintaining the individual dynamic patterns of each axis. The one-dimensional convolutional coding network consists of three one-dimensional convolutional modules. Each convolutional module may include a convolutional layer, a normalization layer, an activation layer, and a pooling layer to map the high-dimensional original temporal waveform into a low-dimensional feature vector. For the first... The axis, whose time characteristics are expressed as , This represents a one-dimensional convolutional encoder with shared weights. In some embodiments, the encoder may also be implemented using a temporal convolutional network (TCN), a recurrent neural network (RNN / LSTM / GRU), a Transformer encoder, or a combination of the above structures.

[0034] The graph attention learning layer uses the temporal features of each axis vibration signal as graph nodes to construct a three-axis adaptive relationship graph. It then determines the relationship score between any two axis vibration signals based on the dot product similarity or the similarity after linear mapping between the temporal features of any two axes in the three-axis adaptive relationship graph. After normalizing each relationship score, the corresponding relationship weight is obtained to construct the three-axis relationship matrix. That is, the obtained three-axis temporal features... , and As graph nodes, a three-axis adaptive relationship graph is constructed. A linear mapping is performed on the temporal features of each axis to obtain the relationship embedding. Then, the dot product similarity or the similarity in the mapping space between any two axes is calculated to obtain the relationship score. .

[0035] In one example, the relationship score can be represented as ,in and Let represent the linear mapping function, and d represent the embedding dimension. The relation scores are normalized to obtain the relation weights. ,in This forms a three-axis relationship matrix. The correlation score is used to characterize the statistical correlation and structural consistency between the three-axis vibration responses under the same mechanical excitation. In some embodiments, the correlation score can also be calculated using cosine similarity, a learnable attention scoring function, or a multilayer perceptron scoring function, and the normalization method can be softmax, sigmoid normalization, or other learnable normalization mechanisms.

[0036] The graph feature fusion layer is based on the triaxial relationship matrix. Graph convolution aggregation is performed on the temporal features of the vibration signal for each axis to weight and converge the structural reference information of the other two axes relative to the current axis, thus obtaining the graph features of the vibration signal for each axis. . It can be represented as , This represents the linear transformation matrix. Furthermore, the convolution time features... With graph features Perform splicing, weighted summation, or gated fusion (or element-wise summation attention fusion or residual fusion) to obtain the fused features for each axis. Subsequently, the fused features from the three directions are concatenated and fully connected to obtain a sample-level diagnostic representation. This design can simultaneously retain both the unique local characteristics of a single axis and the common structural features of multiple axes, while preserving the most significant failure modes.

[0037] The main classification layer will characterize the sample-level diagnostics. The input classification head outputs probability distributions of the triaxial vibration sensor for various different states based on sample-level diagnostic representations, and takes the state with the highest probability as the fault diagnosis result of the triaxial vibration sensor. Specifically, the main classification layer can output probability distributions of the triaxial vibration sensor for six different states: normal state, fixed deviation fault, drift fault, gain fault, jamming fault, and noise fault. The output scores are converted into probabilities using softmax, and the state with the highest probability is taken as the fault diagnosis result of the triaxial vibration sensor.

[0038] In some embodiments, when diagnosing faults in a triaxial vibration sensor, if the probability distribution of the triaxial vibration sensor with respect to various different states determined by the main classification layer has multiple probabilities greater than a preset threshold, a relationship heatmap is generated based on the change in the relationship strength of each axis in the triaxial relationship matrix; then, the abnormal axis in the triaxial vibration sensor is determined based on the relationship heatmap; wherein, if the relationship strength between one axis and the other two axes in the triaxial vibration sensor weakens, it indicates that the relationship between the current axis and the other two axes deviates from the normal triaxial relationship pattern, and is determined to be an abnormal axis.

[0039] In some embodiments, the fault diagnosis model further includes a random masking-assisted training branch. During the training phase, time features or fused features of at least one axis vibration signal are randomly selected for masking processing. The masked features are then recovered using a pre-built auxiliary reconstruction network. An auxiliary classification head is then used to assist in fault diagnosis of the triaxial vibration sensor based on the recovered features. This random masking-assisted training branch enhances the model's robustness to abnormal axis signal loss, contamination, or distortion, enabling the model to perform effective diagnosis using the remaining directions and relational structures even when features in one direction are damaged. In some embodiments, random masking can be performed on one direction, two directions, or local time segments; the auxiliary reconstruction target can be the masked feature or the original time series of the masked axis.

[0040] (3) Model training process: The obtained training set samples are input into the network structure of the above fault diagnosis model in batches. Each batch of samples contains a three-axis time window and a corresponding state category label.

[0041] The temporal feature extraction layer of the weight-sharing CNN is obtained , and The adaptive graph attention learning layer calculates relation scores, relation weights, and the relation matrix. Sample-level diagnostic representations are obtained through graph convolution and feature fusion layers. The main classification branch outputs the probabilities of six states.

[0042] During the training phase, samples are randomly masked, and the masked features are recovered by using the random mask to assist the training branch, thus obtaining the mask reconstruction result; at the same time, the auxiliary classification result can be output through the auxiliary classification head.

[0043] The total loss function is constructed as follows: , Indicates the primary classification loss. This represents the auxiliary classification loss. Indicates the mask reconstruction loss. and The weighting coefficients are as follows: Specifically, the main classification loss function is constructed based on the error between the output of the main classification layer and the true state of the triaxial vibration sensor; the mask reconstruction loss function is constructed based on the error between the features recovered by the auxiliary reconstruction network and the true features before masking; and the auxiliary classification loss function is constructed based on the error between the output of the auxiliary classification head and the true state of the triaxial vibration sensor. Therefore, the main classification loss is used to optimize the six-class state recognition capability, the auxiliary classification loss is used to enhance the discriminative power of intermediate representations, and the mask reconstruction loss is used to improve the model's robustness to local missing values ​​and anomalies.

[0044] An optimizer is used to iteratively update the model parameters until the validation set metrics converge or the number of training epochs reaches a preset value. In practice, the Adam optimizer can be used to perform backpropagation training. During training, the validation set is used to select the learning rate, training epochs, loss weights, and network parameters, and the model parameters with the best validation performance are saved as the final diagnostic model.

[0045] The fault diagnosis model trained using the above methods can be used for fault diagnosis of triaxial vibration sensors.

[0046] In the actual use of the model, it is only necessary to collect the triaxial vibration signal corresponding to the current moment of the device to be diagnosed, and to divide the vibration signal of each axis into sliding window segments according to the preset window length and preset step size to obtain the vibration signal of multiple different time windows of each axis. Then, the vibration signal of multiple different time windows of each axis is subjected to the same mean and standard deviation normalization preprocessing operation as in the model training stage, without the need for additional input of historical normal reference segments.

[0047] The three-axis time window to be diagnosed is input into the trained fault diagnosis model. Sample-level diagnostic representations are obtained through a weight-sharing CNN temporal feature extraction layer, an adaptive graph attention learning layer, and a graph convolution and feature fusion layer. The main classification branch outputs probability values ​​for six states. The class corresponding to the highest probability is used as the sensor state diagnosis result for the current window.

[0048] If multiple categories have high probabilities simultaneously, further analysis can be performed by combining probability distributions, relationship matrices, and relationship heatmaps. A relationship heatmap is generated based on the changes in relationship strength along each axis in the relationship matrix, serving as an auxiliary explanation for abnormal axis directions. If the relationship strength between a certain axis and the other two axes weakens significantly, it indicates that this axis deviates considerably from the normal three-axis relationship pattern within the current window. Figure 4 A schematic diagram of abnormal relationships along the faulty axes is provided. It is understood that the relationship heatmap is used to assist in interpreting relationship anomalies perceived by the model, rather than as a physical measurement for uniquely identifying the faulty axes. In some embodiments, in addition to the relationship heatmap, a numerical table of relationship matrices, relationship change trend curves, or confidence distributions of abnormal axes may also be output.

[0049] The fault diagnosis method for the triaxial vibration sensor of the present invention can be deployed in industrial control computers, edge computing devices, embedded terminals, servers, and cloud-edge collaborative systems.

[0050] The steps of the various methods described above are only for clarity. In practice, they can be combined into one step or some steps can be split into multiple steps. As long as they include the same logical relationship, they are all within the protection scope of this invention. Adding insignificant modifications or introducing insignificant designs to the algorithm or process, without changing the core design of the algorithm and process, are also within the protection scope of this invention.

[0051] Another embodiment of the present invention relates to a fault diagnosis device for a triaxial vibration sensor. The implementation details of the triaxial vibration sensor fault diagnosis device of this embodiment are described below. The following implementation details are provided for ease of understanding and are not essential for implementing this solution. The triaxial vibration sensor fault diagnosis device of this embodiment includes: The signal acquisition module is used to acquire the vibration signals collected by the triaxial vibration sensor along the three axes during the operation of the target device. The relationship construction module is used to extract the temporal features of each axis vibration signal and determine the relationship score between any two axis vibration signals by the dot product similarity or the similarity after linear mapping between the temporal features of any two axes, so as to construct a three-axis relationship matrix. The three-axis relationship matrix is ​​used to characterize the statistical correlation and structural consistency between the three-axis vibration signals of the three-axis vibration sensor. The feature extraction module is used to take the time features of each axis vibration signal as graph nodes, perform graph convolution aggregation on the time features of each axis according to the three-axis relationship matrix, and weight and aggregate the structural reference information of the other two axes to the current axis to obtain the graph features of each axis vibration signal. The graph features of each axis vibration signal are then fused with the time features to obtain the fused features that include the individual features of the corresponding axis vibration signal and the common features of the three axes. The fault diagnosis module is used to diagnose faults in triaxial vibration sensors by using the fusion characteristics of triaxial vibration signals.

[0052] It is not difficult to see that this embodiment is a device embodiment corresponding to the above method embodiments, and this embodiment can be implemented in conjunction with the above method embodiments. The relevant technical details and technical effects mentioned in the above embodiments are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above embodiments.

[0053] It is worth mentioning that all modules involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. Furthermore, to highlight the innovative aspects of this invention, this embodiment does not introduce units that are not closely related to solving the technical problem proposed by this invention; however, this does not mean that other units are absent from this embodiment.

[0054] Another embodiment of the present invention relates to a computer device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the fault diagnosis method for the triaxial vibration sensor in the above embodiments.

[0055] The memory and processor are connected via a bus, which can include any number of interconnecting buses and bridges, connecting various circuits of one or more processors and memories. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which further receives data and transmits it to the processor.

[0056] The processor manages the bus and general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory is used to store data used by the processor during operation.

[0057] Another embodiment of the present invention relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the method embodiments described above.

[0058] That is, those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a device (which may be a microcontroller, chip, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0059] Those skilled in the art will understand that the above embodiments are specific embodiments for implementing the present invention, and in practical applications, various changes can be made to them in form and detail without departing from the spirit and scope of the present invention.

Claims

1. A fault diagnosis method for a triaxial vibration sensor, characterized in that, The method includes: Acquire vibration signals from the three axes of the triaxial vibration sensor during the operation of the target equipment; The temporal features of each axis vibration signal are extracted, and the relationship score between any two axis vibration signals is determined by the dot product similarity or the similarity after linear mapping between the temporal features of any two axes, so as to construct a triaxial relationship matrix. The triaxial relationship matrix is ​​used to characterize the statistical correlation and structural consistency between the triaxial vibration signals of the triaxial vibration sensor. Using the temporal features of each axis vibration signal as graph nodes, graph convolution aggregation is performed on the temporal features of each axis according to the three-axis relationship matrix to weight and aggregate the structural reference information of the other two axes to the current axis, so as to obtain the graph features of each axis vibration signal. The graph features of each axis vibration signal are then fused with the temporal features to obtain the fused features that include the individual features of the corresponding axis vibration signal and the common features of the three axes. Fault diagnosis of triaxial vibration sensors is performed by fusing the characteristics of triaxial vibration signals.

2. The fault diagnosis method for a triaxial vibration sensor according to claim 1, characterized in that, The fault diagnosis results of the triaxial vibration sensor are obtained by inputting the vibration signals collected from the three axes into a trained fault diagnosis model; wherein, the fault diagnosis model includes a temporal feature extraction layer, a graph attention learning layer, a graph feature fusion layer, and a main classification layer; The temporal feature extraction layer uses a three-parameter-shared one-dimensional convolutional coding network to map the vibration signal of each axis to the corresponding temporal features; the one-dimensional convolutional coding network contains three one-dimensional convolutional modules; The graph attention learning layer uses the temporal features of each axis vibration signal as graph nodes to construct a three-axis adaptive relationship graph. The relationship score between any two axis vibration signals is determined by the dot product similarity or the similarity after linear mapping between the temporal features of any two axes in the three-axis adaptive relationship graph. Then, the corresponding relationship weight is obtained by normalizing each relationship score to construct a three-axis relationship matrix. The graph feature fusion layer performs graph convolution aggregation on the temporal features of each axis vibration signal based on the triaxial relationship matrix, so as to weight and converge the structural reference information of the other two axes to the current axis to obtain the graph features of each axis vibration signal. The graph features of each axis vibration signal are then concatenated with the temporal features, weighted summation or gated fusion is performed to obtain the corresponding fused features. Finally, the fused features of the triaxial vibration signals are concatenated and fully connected to obtain the sample-level diagnostic representation. The main classification layer outputs the probability distribution of the triaxial vibration sensor for various different states based on the sample-level diagnostic characterization, and takes the state with the highest probability as the fault diagnosis result of the triaxial vibration sensor.

3. The fault diagnosis method for a triaxial vibration sensor according to claim 2, characterized in that, The fault diagnosis of the triaxial vibration sensor includes: If, based on the probability distribution of the triaxial vibration sensor with respect to various different states, it is determined that there are multiple probabilities greater than a preset threshold, then a relationship heatmap is generated based on the change in the relationship strength of each axis in the triaxial relationship matrix. The abnormal axis in the triaxial vibration sensor is determined based on the relationship heatmap. If the relationship between one axis and the other two axes of the triaxial vibration sensor weakens, it indicates that the relationship between the current axis and the other two axes deviates from the normal triaxial relationship pattern and is determined to be an abnormal axis.

4. The fault diagnosis method for a triaxial vibration sensor according to claim 2, characterized in that, The fault diagnosis model is trained in the following way: At least one axis vibration signal is randomly selected for time feature or fusion feature masking, and the masked features are recovered through a pre-constructed auxiliary reconstruction network. The auxiliary classification head is used to assist in the fault diagnosis of the triaxial vibration sensor based on the recovered features. The main classification loss function is constructed based on the error between the output of the main classification layer and the true state of the triaxial vibration sensor; the mask reconstruction loss function is constructed based on the error between the features recovered by the auxiliary reconstruction network and the true features before masking; and the auxiliary classification loss function is constructed based on the error between the output of the auxiliary classification head and the true state of the triaxial vibration sensor. A fault diagnosis model is trained using the main classification loss function, the mask reconstruction loss function, and the auxiliary classification loss function.

5. The fault diagnosis method for a triaxial vibration sensor according to claim 4, characterized in that, The sample data used during the training of the fault diagnosis model includes samples in several different states: normal state samples, fixed deviation fault samples, drift fault samples, gain fault samples, jammed fault samples, and noise fault samples. The main classification layer outputs the probability distribution of the triaxial vibration sensor for various states such as normal state, fixed deviation fault, drift fault, gain fault, jamming fault and noise fault based on the sample-level diagnostic characterization, and takes the state with the highest probability as the fault diagnosis result of the triaxial vibration sensor.

6. The fault diagnosis method for a triaxial vibration sensor according to claim 5, characterized in that, The samples of the various different states are generated according to the following fault generation rules: Normal state samples are generated from vibration signals collected by a triaxial vibration sensor in a healthy state; fixed deviation fault samples are generated by superimposing a constant bias onto the vibration signal of the target axis of the triaxial vibration sensor; drift fault samples are generated by superimposing a low-frequency trend term that changes over time onto the vibration signal of the target axis of the triaxial vibration sensor; gain fault samples are generated by scaling or amplifying the amplitude of the vibration signal of the target axis of the triaxial vibration sensor; jamming fault samples are generated by replacing the local or overall vibration signal of the target axis of the triaxial vibration sensor with constant or repetitive values; noise fault samples are generated by superimposing random noise and local burst noise onto the vibration signal of the target axis of the triaxial vibration sensor.

7. The fault diagnosis method for a triaxial vibration sensor according to claim 1, characterized in that, Before extracting the time characteristics of the vibration signal for each axis, the method further includes: The vibration signal of each axis is segmented into sliding windows according to a preset window length and a preset step size to obtain vibration signals of multiple different time windows for each axis. Then, the vibration signals of multiple different time windows for each axis are subjected to preprocessing operations of mean removal and standard deviation normalization.

8. A fault diagnosis device for a triaxial vibration sensor, characterized in that, The device includes: The signal acquisition module is used to acquire the vibration signals collected by the triaxial vibration sensor along the three axes during the operation of the target device. The relationship construction module is used to extract the temporal features of each axis vibration signal and determine the relationship score between any two axis vibration signals by the dot product similarity or the similarity after linear mapping between the temporal features of any two axes, so as to construct a three-axis relationship matrix. The three-axis relationship matrix is ​​used to characterize the statistical correlation and structural consistency between the three-axis vibration signals of the three-axis vibration sensor. The feature extraction module is used to take the time features of each axis vibration signal as graph nodes, perform graph convolution aggregation on the time features of each axis according to the three-axis relationship matrix, and weight and aggregate the structural reference information of the other two axes to the current axis to obtain the graph features of each axis vibration signal. The graph features of each axis vibration signal are then fused with the time features to obtain the fused features that include the individual features of the corresponding axis vibration signal and the common features of the three axes. The fault diagnosis module is used to diagnose faults in triaxial vibration sensors by using the fusion characteristics of triaxial vibration signals.

9. A computer device, characterized in that, include: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform a fault diagnosis method for a triaxial vibration sensor as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the fault diagnosis method for the triaxial vibration sensor as described in any one of claims 1 to 7.

Images