Fault diagnosis system and industrial robot state detection and fault diagnosis device

Abstract

The invention provides a fault diagnosis system and an industrial robot state detection and fault diagnosis device, and the fault diagnosis system comprises a signal preprocessing module, a time-frequency feature extraction module, a fault diagnosis module and a classifier, the fault diagnosis module adopts a hybrid refining attention mechanism to carry out weight distribution on original features, spliced enhanced time domain features and frequency domain combination features respectively, important time-frequency features are screened out to carry out weighted fusion, and associated fault features are enriched and highlighted, so that multi-dimensional key fault features of label samples are extracted; and the classifier constructs a mapping relation between the features extracted by the fault diagnosis module and the fault labels by using a full connection layer, classifies the faults, and obtains a fault type diagnosis result. According to the method, the diagnosis precision and efficiency of typical faults such as wear of the reducer and clamping stagnation of the bearing of the industrial robot are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of industrial robot monitoring technology, specifically to a fault diagnosis system, and also to a device comprising the fault diagnosis system and an industrial robot condition detection and fault diagnosis device. Background Technology

[0002] With the development of industrial automation and intelligence, industrial robots are being used more and more widely in manufacturing. When performing various tasks, robots often need to operate for extended periods in complex environments, making their operational status monitoring particularly important.

[0003] Traditional industrial robot fault diagnosis typically requires specialized engineers or technicians to determine the cause of mechanical equipment failure through observation, testing, and analysis. Without sufficient experience, technicians cannot accurately diagnose the machine's fault status in a timely manner for repair, impacting production efficiency. Therefore, fault diagnosis equipment is needed to enable even inexperienced workers to diagnose and handle mechanical faults promptly.

[0004] Existing data acquisition and fault diagnosis equipment has certain limitations in terms of size, weight, power consumption, and data processing capabilities. It cannot simultaneously utilize data from different sensors to combine with advanced artificial intelligence and big data technologies for fault diagnosis, and it cannot meet the current demand for real-time on-site data acquisition and fault diagnosis. Summary of the Invention

[0005] To address the problems in the prior art, the present invention provides a fault diagnosis system, and also provides a fault diagnosis system and an industrial robot condition detection and fault diagnosis device, which aim to solve the power frequency interference problem in current signal analysis and the fault diagnosis challenges in industrial scenarios.

[0006] The fault diagnosis system of this invention includes: Signal preprocessing module: used to acquire the original signal, perform feature enhancement on the original signal to obtain the enhanced time domain signal, and then convert the original signal and the enhanced time domain signal into frequency domain signals respectively and concatenate them to obtain the frequency domain merged signal; Time-frequency feature extraction module: used to extract features from the original signal, enhanced time-domain signal and frequency-domain merged signal through neural convolution, to obtain original features, enhanced time-domain features and frequency-domain merged features. Among them, the original features are further enhanced and concatenated with the enhanced time-domain features to obtain the concatenated enhanced time-domain features. Fault diagnosis module: A hybrid refined attention mechanism is used to assign weights to the original features, the spliced ​​enhanced time-domain features and the frequency-domain merged features respectively, and important time-frequency features are selected and weighted to fuse, enriching and highlighting the associated fault features, thereby extracting key fault features from the label samples in multiple dimensions. Classifier: A fully connected layer is used to construct a mapping relationship between the features extracted by the fault diagnosis module and the fault labels, so as to classify the faults and obtain the fault type diagnosis results.

[0007] Furthermore, the signal processing module includes a first Gramian feature enhancement module, a first Fourier transform module, a second Fourier transform module, and a signal splicing module. The Gramian feature enhancement module receives the original signal as input and outputs an enhanced time-domain signal. The input of the first Fourier transform module is connected to the output of the Gramian feature enhancement module. The input of the second Fourier transform module receives the original signal. The input of the signal splicing module is connected to the outputs of both the first and second Fourier transform modules. The output of the signal splicing module outputs a frequency-domain combined signal.

[0008] Furthermore, the first Gramian feature enhancement module processes the input signal as follows: (1) Convert the collected vibration signals into a two-dimensional signal matrix. The vibration signal includes a series of sample points, and the vibration signal is represented as follows: L is the length of the sample points. , The original one-dimensional vibration signal is segmented and rearranged into a two-dimensional signal matrix X. To facilitate data conversion, the sample length L is set to... ,in, It is an integer, if If the number is even, then the dimension of the original signal matrix constructed is... ,if If the number is odd, then the dimension of the original signal matrix constructed is... ; (2) Calculate the Gramian matrices of the row and column vectors of the two-dimensional signal matrix X. and ; (3) The newly generated Gramian matrix and Treating it as a feature enhancement matrix, multiplying it with the original two-dimensional signal matrix enhances the representation of its internal fault characteristics, resulting in a two-dimensional feature-enhanced signal matrix. : (4) Enhance the two-dimensional feature signal matrix Expanding by row yields a one-dimensional feature enhancement signal. , .

[0009] Furthermore, the time-frequency feature extraction module includes three convolutional layer processing modules, a second Gramian feature enhancement module, and a feature concatenation module. The three inputs of the three convolutional layer processing modules correspond one-to-one with the three output signals of the signal preprocessing module. The first convolutional layer processing module inputs the original signal and outputs the original features. The second convolutional layer processing module inputs the enhanced time-domain signal, and the first convolutional layer processing module inputs the frequency-domain merged signal. The input of the second Gramian feature enhancement module is connected to the output of the first convolutional layer processing module. The input of the feature concatenation module is connected to the outputs of both the second convolutional layer processing module and the second Gramian feature enhancement module.

[0010] Furthermore, the convolutional layer processing module includes a one-dimensional convolutional operation layer, a batch normalization layer, an activation function, and a random dropout layer arranged sequentially according to the signal processing order. The one-dimensional convolutional operation layer is used to initially extract features from the input signal, the batch normalization layer is used to alleviate the gradient diffusion problem and improve the generalization ability of the model, the activation function is used to introduce nonlinearity so that the feature extraction layer can learn and simulate the complex nonlinear relationship between the signal and features, and the random dropout layer is used to discard a portion of neurons during the training iteration process to prevent the model from overfitting due to excessive learning of the training data.

[0011] Furthermore, the fault diagnosis module includes: Channel Attention Refinement Module: Used to obtain channel attention refinement weights; First Refinement Module: Multiply the input features element-wise with the channel attention refinement weights to obtain the channel refined features; Spatial attention refinement module: It is used to divide the channel refined features into first channel refined features and second channel refined features, and then obtain the first spatial attention refinement weight based on the first channel refined features and the second spatial attention refinement weight based on the second channel features. The second refining module: multiply the first channel refined features element by element with the first spatial attention refined weights to obtain the first spatial refined features, and multiply the second channel refined features element by element with the second spatial attention refined weights to obtain the second spatial refined features. Fusion Module: Used to fuse the first space refinement features and the second space refinement features to obtain the final refinement features.

[0012] Furthermore, the processing method of the channel attention refinement module is as follows: Extracting overall feature information from feature channels using an average pooling layer It uses a max pooling layer to extract the most salient feature information from the channels of the input features. An adaptive weighting mechanism is used to aggregate information from both methods, and trainable parameters are set. and During training, it is used to adaptively adjust the feature weights of the features after average pooling and max pooling, and to refine the aggregated features. The input is fed into a convolutional layer for learning, resulting in refined channel attention weights for each channel. .

[0013] Furthermore, the processing method of the spatial attention refinement module is as follows: By comparing the channel attention refinement weights, we determine which channels are important and which are less important, and then use the features of the important channels. and secondary features Separate and aggregate using average pooling and max pooling. and Obtain channel dimension information above and By using convolutional layers to extract features from aggregated channel-dimensional information, important channel features can be obtained. Spatial attention weights and secondary channel characteristics Spatial attention weights The calculation formula is: in, This represents a series of nonlinear computational operations, including batch normalization, ReLU activation function, and sigmoid activation function. This represents a one-dimensional convolution operation with 7 kernels.

[0014] The present invention also provides an industrial robot condition detection and fault diagnosis device, including the fault diagnosis system, a data acquisition system and a condition monitoring system, wherein the data acquisition system is used to collect various sensor parameter information of the industrial robot and then upload it to a host computer; The status monitoring system is used to acquire signals collected by the data acquisition system, construct a multi-dimensional health indicator system through time-domain feature parameters, observe multi-channel signals collected by different sensors in real time, calculate time-domain indicators of different channels, and monitor the time-domain characteristics of the industrial robot in real time.

[0015] Furthermore, the time-domain feature parameters include dimensionless values ​​of spectral kurtosis, peak-to-peak value, kurtosis, waveform factor, and spectral energy. The state monitoring system extracts the dimensionless value with the lowest correlation to rotational speed using the optimal order method and inputs it into the model to obtain the state monitoring results.

[0016] Compared with existing technologies, the beneficial effects of this invention are: by constructing a multi-dimensional health indicator system and combining it with machine learning algorithms to achieve real-time status monitoring, when data anomalies occur, it can quickly diagnose fault types while collecting data, which is fast and accurate.

[0017] Because a faulty bearing generates a unit pulse force, causing high-frequency resonance in the machine, the acquired signal contains periodic pulses. This means the fault data consists of a series of signal bands with similar waveforms. Utilizing this characteristic of vibration signals, they can be reconstructed into a two-dimensional vibration signal matrix. By using a Gramian matrix to enhance the fault characteristics and suppress environmental noise, the signal acquisition accuracy can be effectively improved.

[0018] This invention is compatible with both offline and online diagnostic scenarios. It uses the KSI method to dynamically determine the optimal eigenvalues ​​and optimal order parameters, thereby achieving standardization of vibration characteristics under varying rotational speeds.

[0019] This invention utilizes a signal preprocessing module and a time-frequency feature enhancement module to extract time-frequency enhancement features from the three signal branches of the original signal. In order to enable the model to focus more on key features related to the fault state, a hybrid refined attention mechanism is constructed to assign weights to the features. Important time-frequency features are further selected and weighted to enrich and highlight the associated fault features, thereby fully absorbing the multi-dimensional key fault features of the labeled samples and achieving better robustness.

[0020] This invention significantly improves the diagnostic accuracy and efficiency of typical faults in industrial robots, such as reducer wear and bearing jamming, through a progressive analysis logic of signal characteristics, status indicators, and fault categories. Attached Figure Description

[0021] To more clearly illustrate the solutions in this invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is a block diagram of the fault diagnosis system of the present invention; Figure 2 This is a schematic diagram illustrating the signal feature enhancement effect based on the Gramian matrix. Figure 3 This is a schematic diagram of the convolutional layer processing module structure for the time-frequency feature enhancement module; Figure 4 This is a schematic diagram of the fault diagnosis module structure; Figure 5 This is a schematic diagram of the channel attention refining module structure; Figure 6 This is a schematic diagram of the spatial attention refinement module structure; Figure 7 This is a schematic diagram of the industrial robot condition detection and fault diagnosis device of the present invention. Detailed Implementation

[0023] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings are used to distinguish different objects, not to describe a particular order.

[0024] In this invention, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment to other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this invention can be combined with other embodiments.

[0025] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0026] like Figure 1 As shown, the fault diagnosis system of the present invention includes: a signal preprocessing module, five cascaded time-frequency feature extraction modules, a fault diagnosis module, and a classifier. Preferably, the fault diagnosis module in this example employs a hybrid refined attention mechanism module. The following provides a detailed description of each module.

[0027] The signal preprocessing module is used to acquire the original signal, perform feature enhancement on the original signal to obtain the enhanced time-domain signal, and then convert the original signal and the enhanced time-domain signal into frequency-domain signals respectively and concatenate them to obtain the frequency-domain merged signal.

[0028] The signal processing module in this example includes a first Gramian feature enhancement module, a first Fourier transform module, a second Fourier transform module, and a signal splicing module. The input terminal of the Gramian feature enhancement module receives the original signal and outputs an enhanced time-domain signal. The input terminal of the first Fourier transform module is connected to the output terminal of the Gramian feature enhancement module. The input terminal of the second Fourier transform module receives the original signal. The input terminal of the signal splicing module is connected to the output terminals of the first and second Fourier transform modules, respectively. The output terminal of the signal splicing module outputs a frequency-domain combined signal.

[0029] Because a faulty bearing in an industrial robot generates a unit pulse force, causing high-frequency resonance in the machine, the acquired signal contains periodic pulses. This means the fault data consists of a series of signal bands with similar waveforms. Utilizing this characteristic of vibration signals, they can be reconstructed into a two-dimensional vibration signal matrix. The Gramian matrix can then be used to enhance the fault characteristics and suppress environmental noise.

[0030] A Gramian matrix is ​​a matrix composed of any k vectors in n-dimensional Euclidean space. Each element of the matrix is ​​the inner product of each pair of these vectors. Since the inner product represents the coherence between two vectors, the Gramian matrix, by including the inner products of all vector pairs, can characterize the relationships between these vectors. It can be used to analyze the intrinsic structure of data, calculate the similarity between data points, and capture the features inherent in the data. Given a real matrix , matrix ...

[0031] The signal enhancement method in this example based on the Gramian matrix is ​​as follows: One-dimensional vibration signals collected from mechanical equipment using vibration sensors typically consist of a series of sample points, which can be represented as follows: Its length is One sample point, The original one-dimensional vibration signal is segmented and rearranged into a two-dimensional signal matrix, as shown in the following equation: The dimensions of the original signal matrix depend on the sample length L of the original data. In this example, for ease of data transformation, Set the sample length to ,in, It is an integer. If If the number is even, then the dimension of the original signal matrix constructed is... ,if If the number is odd, then the dimension of the original signal matrix constructed is... .

[0032] For two-dimensional signal matrix Calculate the Gramian matrix of its row vectors and column vectors. and After the inner product operation, the common fault features between the row and column vectors are enhanced, while the randomly occurring noise features are reduced. The correlation of the original internal information features of the matrix is ​​highlighted, which can make the newly generated Gramian matrix more concise. and Treated as a feature enhancement matrix, it is multiplied by the original two-dimensional signal matrix to enhance the representation of its internal fault characteristics, as shown in the following equation: Two-dimensional feature enhancement signal matrix Expanding by row yields a one-dimensional feature enhancement signal. As shown in the following formula: like Figure 2 As shown in the figure, the original simulation signal in this example is shown in subfigure (a), the noise-interference-laden simulation signal is shown in subfigure (b), and the effect of the first Gramian feature enhancement module in enhancing the noise-laden simulation signal is shown in subfigure (c). It can be seen that the present invention can effectively suppress noise interference and effectively enhance the features.

[0033] Due to the signal feature enhancement effect of the Gramian matrix, it can be incorporated into deep learning network models to extract enhanced signals from the original signal and strengthen the feature extraction capabilities of the deep learning network model. The signal preprocessing module of this invention uses one-dimensional vibration data of the original signal as input and the Gramian matrix as a feature enhancement module to enhance the signal features of the original data, forming a feature-enhanced signal branch. Simultaneously, fast Fourier transforms are performed on the original signal and its enhanced signal, and these are combined to form a frequency domain signal branch. This branch, along with the original signal and its enhanced signal, forms three input signal branches, which serve as inputs to the network model for the next step of deep feature extraction. This allows the model to simultaneously extract time-domain and frequency-domain features of the original signal and its enhanced signal, enabling observation of fault features from different perspectives and providing sufficient data for obtaining effective fault features characterizing fault modes.

[0034] like Figure 1 and Figure 3As shown, the time-frequency feature extraction module in this example is used to extract features from the original signal, the enhanced time-domain signal, and the frequency-domain merged signal through neural convolution, to obtain the original features, the enhanced time-domain features, and the frequency-domain merged features. The original features are further enhanced and concatenated with the enhanced time-domain features to obtain the concatenated enhanced time-domain features. Of course, the number of time-frequency feature extraction modules can also be set according to the situation.

[0035] The time-frequency feature extraction module includes three convolutional layer processing modules, a second Gramian feature enhancement module, and a feature concatenation module. The three inputs of the three convolutional layer processing modules correspond one-to-one with the three output signals of the signal preprocessing module. The first convolutional layer processing module receives the original signal as input and outputs the original features. The second convolutional layer processing module receives the enhanced time-domain signal as input, and the first convolutional layer processing module receives the frequency-domain merged signal as input. The input of the second Gramian feature enhancement module is connected to the output of the first convolutional layer processing module. The input of the feature concatenation module is connected to the outputs of both the second convolutional layer processing module and the second Gramian feature enhancement module. In this example, the second Gramian feature enhancement module operates on the same principle as the first Gramian feature enhancement module, and will not be described further here.

[0036] In this example, to extract the potential relationship between signal samples and health status from three signal branches, a Time-Frequency Feature Enhancement Module (TFFEM) is designed as a deep learning network module for feature extraction. The convolutional layer processing module includes a one-dimensional convolutional layer, a batch normalization layer, an activation function, and a random dropout layer, arranged sequentially according to the signal processing order. The convolutional layer is the feature extraction layer, and its components are as follows: Figure 3 As shown in the diagram. The Conv1D layer (one-dimensional convolutional layer) is used to initially extract features from the input signal, and the BN layer (batch normalization layer) is used to alleviate gradient spread and improve the model's generalization ability. ReLU is an activation function used to introduce nonlinearity, enabling the feature extraction layer to learn and simulate the complex nonlinear relationship between the signal and features. The Dropout layer is used to discard a portion of neurons during training iterations to prevent overfitting due to excessive learning of the training data. To ensure that fault features in the signal are fully extracted, the fault features extracted from the original signal are also enhanced using a Gramian-based feature enhancement method, merging with the fault features extracted from the enhanced signal by the feature extraction layer as an enhanced feature extraction branch. This, along with the original signal feature extraction branch and the frequency domain signal feature extraction branch, constitutes the time-frequency feature extraction module.

[0037] like Figure 1 and Figure 4 As shown, the fault diagnosis module in this example uses a hybrid refined attention mechanism to assign weights to the original features, the spliced ​​enhanced time-domain features, and the frequency-domain merged features, respectively. It further selects important time-frequency features for weighted fusion, enriches and highlights the associated fault features, and thus extracts key fault features from the label samples in multiple dimensions.

[0038] The classifier uses a fully connected layer to construct a mapping relationship between the features extracted by the fault diagnosis module and the fault labels, classifies the faults, and obtains the fault type diagnosis results.

[0039] The fault diagnosis module includes: Channel Attention Refinement Module: Used to obtain channel attention refinement weights; First Refinement Module: Multiply the input features element-wise with the channel attention refinement weights to obtain the channel refined features; Spatial attention refinement module: It is used to divide the channel refined features into first channel refined features and second channel refined features, and then obtain the first spatial attention refinement weight based on the first channel refined features and the second spatial attention refinement weight based on the second channel features. The second refining module: multiply the first channel refined features element by element with the first spatial attention refined weights to obtain the first spatial refined features, and multiply the second channel refined features element by element with the second spatial attention refined weights to obtain the second spatial refined features. Fusion Module: Used to fuse the first space refinement features and the second space refinement features to obtain the final refinement features.

[0040] like Figure 5 As shown, in order to measure the importance of features between channels and further assign importance weights, this example compresses the spatial dimension while keeping the channel dimension unchanged, and constructs a channel attention refining module.

[0041] Furthermore, the processing method of the channel attention refinement module is as follows: Extracting overall feature information from feature channels using an average pooling layer It uses a max pooling layer to extract the most salient feature information from the channels of the input features. An adaptive weighting mechanism is used to aggregate information from both methods, and trainable parameters are set. and During training, this is used to adaptively adjust the feature weights of the features after average pooling and max pooling, thereby refining and extracting channel features more effectively. The aggregated refined features... The input is fed into a convolutional layer for learning, resulting in refined channel attention weights for each channel. .

[0042] Acquire channel refinement features The specific operational equation is as follows:

[0043] Obtain the channel importance weights of features The process is as follows: in, This represents the sigmoid activation function. Indicates the number of convolution kernels as One-dimensional convolution operation, This indicates average pooling calculation. This indicates max pooling computation. To better capture cross-channel features, the number of convolutional kernels is adaptively adjusted based on the number of channels. . The specific calculation process is shown in the formula below, where, and These are hyperparameters, typically set to 2 and 1 respectively.

[0044] like Figure 6 As shown in the example, in order to measure which features in the channel are more important, a spatial attention refining module is constructed to process the features refined by the channel attention mechanism, compressing the channel dimension while keeping the spatial dimension unchanged.

[0045] Since each channel has a different level of importance for feature refinement, in order to enhance the role of relevant features from important channels and reduce the interference from irrelevant features from less important channels, we compare the channel attention refinement weights to determine which channels are important and which are less important.

[0046] The processing method for the spatial attention refinement module in this example is as follows: First, distinguish between important and secondary important channels, and then mask the features of the important and secondary important channels respectively, thus masking the features of the important channels. and secondary features In this example, the ratio of the number of items separated is 1:1.

[0047] Secondly, aggregation using average pooling and max pooling. and Obtain channel dimension information above and By using convolutional layers to extract features from aggregated channel-dimensional information, important channel features can be obtained. Spatial attention weights and secondary channel characteristics Spatial attention weights The calculation formula is: in, This represents a series of nonlinear computational operations, including batch normalization, ReLU activation function, and sigmoid activation function. This represents a one-dimensional convolution operation with 7 kernels.

[0048] Finally, the important channel features Spatial attention weights and secondary channel characteristics Spatial attention weights The features are multiplied and combined with the corresponding channel features to obtain the spatial attention-refined features.

[0049] like Figure 4 As shown, the working principle of the hybrid attention mechanism module, which combines the channel attention refinement module and the spatial attention refinement module to construct a hybrid refined attention mechanism, is as follows: For input features To ensure the model focuses on key features in the original feature tensor channels, the hybrid refined attention mechanism module first extracts a channel attention refined weight using the channel attention mechanism. This refined channel feature is then multiplied by the original feature tensor, and attention weights are assigned to the original feature tensor to obtain the channel refined feature. The channel refined feature is then separated into... and Two sets of features are used. Spatial attention is employed to refine the two sets of features after separation, generating refined spatial features. Finally, the two sets of refined spatial features are added and fused to obtain the final features. The specific process is shown in the following expression: Among them, symbols Represents element-wise multiplication, symbol This indicates element-wise addition. This is the feature refinement result of assigning channel attention refinement weights to the features. This is the result of assigning spatial attention weights to the refined features of the channels. It is the final refined feature after channel attention and spatial attention weight allocation.

[0050] In summary, this invention obtains three time-frequency branch signals of the original signal through signal processing, fully explores the hidden nonlinear relationship between fault data and fault features using five time-frequency feature enhancement modules, refines features using a hybrid refined attention mechanism to obtain importance weights for different features, enabling the model to focus more on key features related to fault information, and finally uses a fully connected layer to construct the mapping relationship between the extracted features and fault labels, thereby accurately classifying fault types with high classification accuracy and good robustness.

[0051] like Figure 7 As shown, the present invention also provides an industrial robot condition detection and fault diagnosis device, including the fault diagnosis system, a data acquisition system and a condition monitoring system, wherein the data acquisition system is used to collect various sensor parameter information of the industrial robot and then upload it to a host computer; The status monitoring system is used to acquire signals collected by the data acquisition system, construct a multi-dimensional health indicator system through time-domain feature parameters, observe multi-channel signals collected by different sensors in real time, calculate time-domain indicators of different channels, and monitor the time-domain characteristics of the industrial robot in real time.

[0052] Preferably, in this example, before sending the acquired signal to the status monitoring system, a signal processing step is included to improve the signal quality and thus better monitor the status and fault type analysis of the industrial robot.

[0053] This example acquires multi-channel features of time-domain signals, constructs a multi-dimensional health indicator system using time-domain feature parameters (such as dimensionless values ​​like spectral kurtosis, peak-to-peak value, kurtosis, waveform factor, and spectral energy), and combines this with machine learning algorithms to achieve real-time status warnings.

[0054] The signal processing method in this example is feature normalization: normalized features of faulty and healthy bearings are obtained and input into the Random Forest model as a training set. Then, the normalized features of the test bearings are input into the model as predictions. When the data type of the test bearing is closer to that of a healthy bearing, the model result is 0, and otherwise it is 1.

[0055] Preferably, in this example, the dimensionless value with the lowest correlation to rotational speed is extracted using the KSI method (optimal order method) and input into the model to obtain the condition monitoring results. The specific steps include: (1) Feature extraction: Obtain multiple features Ki and rotational speed signal V. In this example, the features Ki selected are: spectral kurtosis, root mean square in time domain, peak-to-peak value, root mean square in frequency domain, spectral energy, etc.

[0056] (2) Set the order search parameter n, with the range of n being (0.1, 10), and then calculate the normalized features. ; (3) Order optimization: after obtaining normalized features Next, each normalized feature is calculated. The Kendall τ-b correlation coefficient between Ki / V and rotational speed V is used to analyze the relationship between the characteristics and rotational speed. The optimal order n is searched to minimize the correlation between Ki / V^n and rotational speed. The order n* with the lowest correlation coefficient is selected as the optimal order.

[0057] (4) Obtaining normalized features : Using the optimal order n*, we obtain the normalized features The calculation formula is: ， Where ki represents spectral kurtosis, root mean square, peak-to-peak value, root mean square in the frequency domain, spectral energy, etc., and Kendall's τ-b correlation coefficient ranges from -1 to 1. , This represents the computational cost of the formula.

[0058] In a preferred embodiment of the present invention, the industrial robot condition detection and fault diagnosis device is integrated into a portable equipment fault diagnosis instrument. Each system is installed into the equipment fault diagnosis instrument through built-in algorithms, thereby facilitating monitoring and diagnosis. In this example, the collected time-domain signal is input into the equipment fault diagnosis instrument. The instrument performs condition monitoring and fault diagnosis on the bearing through built-in algorithms, and finally provides feedback to the user on the type of fault.

[0059] This invention significantly improves the diagnostic accuracy and efficiency of typical faults in industrial robots, such as reducer wear and bearing jamming, through a progressive analysis logic of signal characteristics, status indicators, and fault categories.

[0060] As can be seen from the above, compared with the prior art, the present invention has the following innovative points: 1. By constructing a multi-dimensional health indicator system and combining it with machine learning algorithms, real-time status monitoring can be achieved. When data anomalies occur, the system can quickly diagnose the fault type while collecting data, which is fast and accurate.

[0061] 2. By using the Gramian matrix to enhance fault characteristics and suppress environmental noise, the accuracy of signal acquisition can be effectively improved.

[0062] 3. This invention is compatible with both offline and online diagnostic scenarios. It uses the KSI method to dynamically determine the optimal eigenvalues ​​and optimal order parameters, thereby achieving standardization of vibration characteristics under varying rotational speeds.

[0063] 4. This invention utilizes a signal preprocessing module and a time-frequency feature enhancement module to extract the time-frequency enhancement features of the three signal branches of the original signal. In order to enable the model to focus more on the key features related to the fault state, a hybrid refined attention mechanism is constructed to assign weights to the features, further enriching and highlighting the associated fault features, thereby fully absorbing the multi-dimensional key fault features of the labeled samples and achieving better robustness.

[0064] 5. This invention significantly improves the diagnostic accuracy and efficiency of typical faults such as reducer wear and bearing jamming in industrial robots through a progressive analysis logic of signal characteristics, status indicators, and fault categories.

[0065] The specific embodiments described above are preferred embodiments of the present invention and are not intended to limit the specific scope of the present invention. The scope of the present invention includes, but is not limited to, these specific embodiments. All equivalent changes made in accordance with the present invention are within the protection scope of the present invention.

Claims

1. A fault diagnosis system, characterized in that, include: Signal preprocessing module: used to acquire the original signal, perform feature enhancement on the original signal to obtain the enhanced time domain signal, and then convert the original signal and the enhanced time domain signal into frequency domain signals respectively and concatenate them to obtain the frequency domain merged signal; Time-frequency feature extraction module: used to extract features from the original signal, enhanced time-domain signal and frequency-domain merged signal through neural convolution, to obtain original features, enhanced time-domain features and frequency-domain merged features. Among them, the original features are further enhanced and concatenated with the enhanced time-domain features to obtain the concatenated enhanced time-domain features. Fault diagnosis module: A hybrid refined attention mechanism is used to assign weights to the original features, the spliced ​​enhanced time-domain features and the frequency-domain merged features respectively, and important time-frequency features are selected and weighted to fuse, enriching and highlighting the associated fault features, thereby extracting key fault features from the label samples in multiple dimensions. Classifier: A fully connected layer is used to construct a mapping relationship between the features extracted by the fault diagnosis module and the fault labels, so as to classify the faults and obtain the fault type diagnosis results.

2. The fault diagnosis system according to claim 1, characterized in that: The signal processing module includes a first Gramian feature enhancement module, a first Fourier transform module, a second Fourier transform module, and a signal splicing module. The input terminal of the Gramian feature enhancement module receives the original signal and outputs an enhanced time-domain signal. The input terminal of the first Fourier transform module is connected to the output terminal of the Gramian feature enhancement module. The input terminal of the second Fourier transform module receives the original signal. The input terminal of the signal splicing module is connected to the output terminals of the first Fourier transform module and the second Fourier transform module, respectively. The output terminal of the signal splicing module outputs a frequency-domain combined signal.

3. The fault diagnosis system according to claim 2, characterized in that: The first Gramian feature enhancement module processes the input signal as follows: (1) Convert the collected vibration signals into a two-dimensional signal matrix. The vibration signal includes a series of sample points, and the vibration signal is represented as follows: L is the length of the sample points. , The original one-dimensional vibration signal is segmented and rearranged into a two-dimensional signal matrix X. To facilitate data conversion, the sample length L is set to... ,in, It is an integer, if If the number is even, then the dimension of the original signal matrix constructed is... ,if If the number is odd, then the dimension of the original signal matrix constructed is... ; (2) Calculate the Gramian matrices of the row and column vectors of the two-dimensional signal matrix X. and ; (3) The newly generated Gramian matrix and Treating it as a feature enhancement matrix, multiplying it with the original two-dimensional signal matrix enhances the representation of its internal fault characteristics, resulting in a two-dimensional feature-enhanced signal matrix. : (4) Enhance the two-dimensional feature signal matrix Expanding by row yields a one-dimensional feature enhancement signal. , 。 4. The fault diagnosis system according to claim 3, characterized in that: The time-frequency feature extraction module includes three convolutional layer processing modules, a second Gramian feature enhancement module, and a feature concatenation module. The three inputs of the three convolutional layer processing modules correspond one-to-one with the three output signals of the signal preprocessing module. The first convolutional layer processing module receives the original signal as input and outputs the original features. The second convolutional layer processing module receives the enhanced time-domain signal as input, and the first convolutional layer processing module receives the frequency-domain merged signal as input. The input of the second Gramian feature enhancement module is connected to the output of the first convolutional layer processing module. The input of the feature concatenation module is connected to the outputs of both the second convolutional layer processing module and the second Gramian feature enhancement module.

5. The fault diagnosis system according to claim 4, characterized in that: The convolutional layer processing module includes a one-dimensional convolutional operation layer, a batch normalization layer, an activation function, and a random dropout layer arranged sequentially according to the signal processing order. The one-dimensional convolutional operation layer is used to initially extract features from the input signal. The batch normalization layer is used to alleviate the gradient diffusion problem and improve the generalization ability of the model. The activation function is used to introduce nonlinearity so that the feature extraction layer can learn and simulate the complex nonlinear relationship between the signal and features. The random dropout layer is used to discard a portion of neurons during the training iteration process to prevent the model from overfitting due to excessive learning of the training data.

6. The fault diagnosis system according to any one of claims 1-5, characterized in that: The fault diagnosis module includes: Channel Attention Refinement Module: Used to obtain channel attention refinement weights; First Refinement Module: Multiply the input features element-wise with the channel attention refinement weights to obtain the channel refined features; Spatial attention refinement module: It is used to divide the channel refined features into first channel refined features and second channel refined features, and then obtain the first spatial attention refinement weight based on the first channel refined features and the second spatial attention refinement weight based on the second channel features. The second refining module: multiply the first channel refined features element by element with the first spatial attention refined weights to obtain the first spatial refined features, and multiply the second channel refined features element by element with the second spatial attention refined weights to obtain the second spatial refined features. Fusion Module: Used to fuse the first space refinement features and the second space refinement features to obtain the final refinement features.

7. The fault diagnosis system according to claim 6, characterized in that: The processing method of the channel attention refinement module is as follows: Extracting overall feature information from feature channels using an average pooling layer It uses a max pooling layer to extract the most salient feature information from the channels of the input features. An adaptive weighting mechanism is used to aggregate information from both methods, and trainable parameters are set. and During training, it is used to adaptively adjust the feature weights of the features after average pooling and max pooling, and to refine the aggregated features. The input is fed into a convolutional layer for learning, resulting in refined channel attention weights for each channel. .

8. The fault diagnosis system according to claim 7, characterized in that: The processing method of the spatial attention refinement module is as follows: By comparing the channel attention refinement weights, we determine which channels are important and which are less important, and then use the features of the important channels. and secondary important features Separate and aggregate using average pooling and max pooling. and Obtain channel dimension information above and By using convolutional layers to extract features from aggregated channel-dimensional information, important channel features can be obtained. Spatial attention weights and secondary channel characteristics Spatial attention weights The calculation formula is: in, This represents a series of nonlinear computational operations, including batch normalization, ReLU activation function, and sigmoid activation function. This represents a one-dimensional convolution operation with 7 kernels.

9. An industrial robot condition detection and fault diagnosis device, comprising the fault diagnosis system according to any one of claims 1-8, characterized in that: It also includes a data acquisition system and a status monitoring system, wherein the data acquisition system is used to collect various sensor parameter information of the industrial robot and then upload it to the host computer; The status monitoring system is used to acquire signals collected by the data acquisition system, construct a multi-dimensional health indicator system through time-domain feature parameters, observe multi-channel signals collected by different sensors in real time, calculate time-domain indicators of different channels, and monitor the time-domain characteristics of the industrial robot in real time.

10. The industrial robot condition detection and fault diagnosis device according to claim 9, characterized in that: The time-domain feature parameters include dimensionless values ​​of spectral kurtosis, peak-to-peak value, kurtosis, waveform factor, and spectral energy. The state monitoring system extracts the dimensionless value with the lowest correlation to rotational speed using the optimal order method and inputs it into the model to obtain the state monitoring results.

Images