Fault diagnosis method for harmonic reducer based on multi-source heterogeneous data fusion

Abstract

The present invention discloses a harmonic reducer fault diagnosis method based on multi-source heterogeneous data fusion, which relates to the field of collaborative robot joint fault diagnosis. In order to solve the problem of low diagnosis rate of existing harmonic reducer fault diagnosis methods, the present invention designs a harmonic reducer fault diagnosis method based on multi-source heterogeneous data fusion. By collecting the original data of multiple sensors such as vibration, torque and temperature of the harmonic reducer under different fault modes, the one-dimensional time domain signals of multiple sensors are converted into two-dimensional time-frequency images through continuous wavelet transform, and a fault feature time-frequency diagram dataset of the harmonic reducer is constructed. Feature selection and noise suppression are performed through the residual shrinkage unit in the DRSN-SE algorithm, and the feature representation capability is enhanced by the SE-Net attention mechanism. The features of multi-source heterogeneous data are fused to obtain the final fused features and output the fault diagnosis results through the classification head. The present invention can effectively solve the problem of poor reliability of single fault data and realize high-precision fault diagnosis and classification identification of harmonic reducers.

Description

Technical Field

[0001] The present invention belongs to the field of big data learning models, and specifically relates to a harmonic reducer fault diagnosis method based on multi-source heterogeneous data fusion. Background Art

[0002] As a core component of robot joints, the operating status of intelligent harmonic reducers has a significant impact on the performance and lifespan of the entire system. Due to factors such as complex working environments and variable loads, harmonic reducers are prone to faults such as flexible wheel wear and fracture, rigid wheel wear, and bearing pitting, which can lead to serious consequences such as personal injury, death, and economic losses due to robot failures. At present, fault detection and performance testing of harmonic reducers are mainly carried out by workshop technicians using vibration analyzers to complete comparative tests for troubleshooting. This is inefficient and cannot be integrated into the robot joint system. Research on harmonic reducer fault diagnosis is still very scarce, and it is impossible to monitor potential faults in real time and reduce the losses caused by faults. Therefore, using the signals generated during the operation of the harmonic reducer to perform intelligent fault diagnosis and analysis is of great guiding significance for improving the working performance of the robot.

[0003] Harmonic reducers are often affected by the coupling of external systems such as motors and loads during operation. Factors such as noise and harmonic signals make it difficult for the characteristic information of a single signal to reflect the actual fault state of the harmonic reducer, significantly reducing the recognition effectiveness of the fault diagnosis system. Therefore, it is necessary to integrate information from multiple different types of sensors to achieve complementary features from multi-source heterogeneous data and improve the robustness of the fault diagnosis system. Summary of the Invention

[0004] Aiming at the problems of weak characteristic information of single sensor signal of harmonic reducer, difficulty in early fault identification, unclear multi-source heterogeneous data fusion mechanism, and missing data sets of different fault modes, the present invention proposes a harmonic reducer fault diagnosis method based on multi-source heterogeneous data fusion.

[0005] The specific technical solution adopted by the present invention includes the following steps:

[0006] A harmonic reducer fault diagnosis method based on multi-source heterogeneous data fusion includes the following steps:

[0007] S1. Data sensors placed on the harmonic reducer are used to collect one-dimensional time domain signals of vibration, torque, and temperature of the harmonic reducer under different fault modes. The three one-dimensional time domain signals collected under each fault mode constitute a data sample in the multi-source heterogeneous dataset.

[0008] S2. Preprocess each data sample in the multi-source heterogeneous data, convert the one-dimensional time domain signals of vibration, torque, and temperature into two-dimensional time-frequency domain feature maps using the continuous wavelet transform method, thereby converting each data sample into a training sample and constructing a training sample set with fault mode labels;

[0009] S3. Using the training sample set to train a harmonic reducer fault diagnosis model; the input of the harmonic reducer fault diagnosis model is a two-dimensional time-frequency domain feature map of three signals: vibration, torque, and temperature. The two-dimensional time-frequency domain feature maps of the three signals are first subjected to noise suppression and feature selection by a deep residual shrinkage network, and then the important features are enhanced by the SE-Net module. Finally, the three enhanced features are fused at the feature level and a prediction head is used to output a fault prediction label for the harmonic reducer;

[0010] S4. For a harmonic reducer with an unknown fault mode, obtain the one-dimensional time domain signals of vibration, torque and temperature and convert them into two-dimensional time-frequency domain feature maps respectively through the continuous wavelet transform method, and then input them into the trained harmonic reducer fault diagnosis model to predict the fault mode of the harmonic reducer.

[0011] Preferably, the data sensors arranged on the harmonic reducer include an acceleration sensor for detecting vibration signals, a torque sensor for detecting the output torque of the harmonic reducer, and a temperature sensor for detecting the operating temperature of the harmonic reducer.

[0012] Preferably, the vibration sensor is adsorbed on the outer surface of the rigid wheel by magnetic attraction; the torque sensor is installed at the output end of the harmonic reducer; the temperature sensor uses a temperature probe and is attached to the surface of the harmonic reducer to measure the real-time operating temperature.

[0013] Preferably, the failure modes covered by the data samples in the multi-source heterogeneous data set include six types: no fault, rigid wheel wear fault, flexspline wear fault, flexspline pitting fault, flexspline crack fault, and bearing pitting fault.

[0014] Preferably, the one-dimensional time domain signal collected by each data sensor needs to be resampled and normalized to the minimum and maximum value, and then divided into signal segments of fixed length and converted into a two-dimensional time-frequency feature map through continuous wavelet transform. The two-dimensional time-frequency feature map obtained by continuous wavelet transform is resampled by interpolation to meet the input requirements of the harmonic reducer fault diagnosis model.

[0015] Preferably, the harmonic reducer fault diagnosis model includes three branches with the same structure, each sensor signal corresponds to a branch, and each branch is composed of an input layer, two convolution layers, a maximum pooling layer, three two-dimensional residual shrinkage unit modules, a global average pooling layer and an SE-Net module cascaded in sequence. The input layer of the first branch is used to input the two-dimensional time-frequency domain feature map of the vibration signal, the input layer of the second branch is used to input the two-dimensional time-frequency domain feature map of the torque signal, and the input layer of the third branch is used to input the two-dimensional time-frequency domain feature map of the harmonic temperature signal. After passing through the SE-Net module, each of the three branches obtains an enhanced feature; the enhanced features output by the three branches are fused through the feature fusion layer, and then pass through the prediction head composed of a fully connected layer and a Softmax layer to output the classification label of the fault mode.

[0016] Preferably, in each of the two-dimensional residual shrinkage unit modules, the input feature map of the module first passes through two two-dimensional convolutional layers to obtain an intermediate feature map, and the intermediate feature map is subjected to absolute value operation and global average pooling to obtain a feature vector Z; the feature vector Z passes through two fully connected layers and is output by the Sigmoid function to obtain the scaling parameter a of each channel, all channels of the feature vector Z are scaled by their respective scaling parameters to obtain a threshold vector, and the intermediate feature map is soft-thresholded using the threshold vector to eliminate noise-related information in the input feature map; finally, the feature map after soft-threshold processing is connected to the residual of the input feature map and passed to the next cascade network layer through the ReLU function.

[0017] Preferably, in each branch of the harmonic reducer fault diagnosis model, the size of the two-time-frequency feature map input by the input layer is 3×256×256; the filter sizes of the two convolutional layers are 3 and 16 respectively, the convolution kernel size is 3×3, and the stride is 2; the convolution kernel size of the maximum pooling layer is 2×2, and the stride is 2; the size of the three filters in the first residual contraction unit is 32, the convolution kernel size is 3×3, and the stride is 1; the size of the three filters in the second residual contraction unit is 48, the convolution kernel size is 3×3, and the stride is 1; the size of the three filters in the third residual contraction unit is 64, the convolution kernel size is 3×3, and the stride is 1; the feature map size of the input global average pooling layer is 64×4×4, and the output size of the global average pooling layer is a tensor of 64×1×1.

[0018] Preferably, in the feature fusion layer, the three enhanced features are fused at the feature level through a feature splicing operation.

[0019] Preferably, when the harmonic reducer fault diagnosis model is trained using the training sample set, it is necessary to pre-divide it into a training set and a validation set according to a preset ratio.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] 1) The present invention establishes an intelligent fault diagnosis data set for harmonic reducers by collecting multi-source heterogeneous data such as vibration, torque, and temperature of harmonic reducers under different fault modes. This eliminates the need for technicians to manually troubleshoot fault modes, reduces the impact of external systems such as motors and loads on fault diagnosis, and improves the robustness of the harmonic reducer fault diagnosis system.

[0022] 2) Multi-source heterogeneous time domain datasets under different fault modes are converted into time-frequency domain feature maps through continuous wavelet transform. Based on the DRSN-SE model, the noise of the original signal feature maps in multiple branches is suppressed, and feature enhancement at different levels is achieved, thereby improving the classification and recognition performance of the fault system.

[0023] 3) The DRSN-SE model for multi-source heterogeneous data fusion can be integrated into the online monitoring system of robot joints to monitor potential faults of harmonic reducers in real time and reduce the losses caused by faults.

[0024] The above description briefly summarizes the technical solution of the present invention. In order to make the features, solutions and advantages of the present invention easier to understand and use, the following detailed description is provided through specific implementation plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 The figure is a flowchart of the steps of a harmonic reducer fault diagnosis method based on multi-source heterogeneous data fusion.

[0026] Figure 2 Schematic diagram of a data acquisition system in an embodiment of the present invention.

[0027] Figure 3 This is a DRSN-SE fault diagnosis model based on multi-source heterogeneous data fusion in an embodiment of the present invention.

[0028] Figure 4 2D-RSUB unit structure diagram in an embodiment of the present invention.

[0029] Explanation of symbols in the figure: driving servo motor 1, first encoder 2, input torque sensor 3, vibration sensor 4, temperature sensor 5, harmonic reducer 6, output torque sensor 7, second encoder 8, magnetic powder brake 9. DETAILED DESCRIPTION

[0030] In order to make the above-mentioned objects, features and advantages of the present invention more clearly understood, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. The technical features in the various embodiments of the present invention can be combined accordingly without conflicting with each other.

[0031] In the description of the present invention, it should be understood that when an element is considered to be "connected" to another element, it can be directly connected to the other element or indirectly connected, that is, there are intermediate elements. On the contrary, when an element is said to be "directly" connected to another element, there are no intermediate elements.

[0032] In the description of the present invention, it should be understood that the terms "first" and "second" are used solely for descriptive purposes and are not to be construed as indicating or implying relative importance or implicitly specifying the number of technical features being described. Therefore, features defined as "first" or "second" may explicitly or implicitly include at least one of such features.

[0033] In a preferred embodiment of the present invention, a harmonic reducer fault diagnosis method based on multi-source heterogeneous data fusion is provided. Figure 1 As shown, the specific steps include S1 to S4. The specific implementation of each step is described in detail below.

[0034] S1. Through the data sensors arranged on the harmonic reducer, the vibration, torque and temperature signals of the harmonic reducer under different fault modes are collected. The signals collected by different data sensors are all one-dimensional time domain signals under one-dimensional time series.

[0035] The specific form of the above-mentioned data sensor can be adjusted according to actual needs. In the present invention, the data sensor arranged on the harmonic reducer may include an acceleration sensor for detecting vibration signals, a torque sensor for detecting the output torque of the harmonic reducer, and a temperature sensor for detecting the operating temperature of the harmonic reducer.

[0036] The specific arrangement position of the above data sensors can also be adjusted according to actual needs. In order to facilitate accurate data collection, in the embodiment of the present invention, it is necessary to build a harmonic reducer multi-sensor data acquisition test platform to obtain the one-dimensional time domain signals of the three signals of the above harmonic reducer under different fault modes. Figure 2As shown, the basic components of the test platform include a drive servo motor 1, a first encoder 2, an input torque sensor 3, a vibration sensor 4, a temperature sensor 5, a harmonic reducer 6, an output torque sensor 7, a second encoder 8, and a magnetic powder brake 9. From left to right, they are: drive servo motor 1, first encoder 2, input torque sensor 3, harmonic reducer 6, output torque sensor 7, second encoder 8, and magnetic powder brake 9. In this embodiment of the present invention, the temperature sensor 5 uses an acceleration sensor perpendicular to the surface of the harmonic reducer's rigid wheel to collect vibration signals during operation of the harmonic reducer under different fault modes. The temperature sensor 5 uses a K-type thermocouple with a temperature probe attached to the surface of the harmonic reducer 6 to measure the temperature. The input torque sensor 3 and the output torque sensor 7 can capture real-time torque changes of the harmonic reducer using torque sensors at the input and output ends, respectively. In this embodiment of the present invention, the output torque sensor 7, installed at the output end of the harmonic reducer 6, records the one-dimensional time domain signal of the torque. The speed of the harmonic reducer 6 on the test platform is set to 2400 rpm, and the load is set to 80 Nm.

[0037] Based on this test platform, we can simulate harmonic reducers with different fault modes and then collect one-dimensional time domain signals of three different signals. The three one-dimensional time domain signals collected under each fault mode constitute a data sample in the multi-source heterogeneous data set, thereby forming multi-source heterogeneous data and used to construct model training samples.

[0038] In this embodiment of the present invention, the data samples in the multi-source heterogeneous dataset cover six types of failure modes: no fault (i.e., normal operation), rigid wheel wear fault, flexspline wear fault, flexspline pitting fault, flexspline crack fault, and bearing pitting fault. Bearing pitting faults are classified as inner and outer race faults and are processed by electrical discharge cutting, with a width of 1 mm, a length of 3 mm, and a depth of 1 mm.

[0039] Under six different fault modes, the corresponding vibration signals, torque signals, and temperature signals are collected, and the fault diagrams of the harmonic reducer in the six fault modes are recorded. During the sampling process, the data acquisition card acquisition frequency of the vibration signal is set to 25600Hz, the sampling frequency of the torque sensor is set to 2560Hz, and the temperature sensor is 10Hz. The sampling time is 10 minutes, and a sliding window of 12 seconds is selected. The signals collected by the torque sensor and the temperature sensor are resampled, and the number of sampling points of the vibration signal is unified to 307200 points.

[0040] S2: Preprocess each data sample in the multi-source heterogeneous data, convert the one-dimensional time domain signals of vibration, torque and temperature into two-dimensional time-frequency domain feature maps through the continuous wavelet transform method, thereby converting each data sample into a training sample and constructing a training sample set with fault mode labels.

[0041] The one-dimensional time-domain signals collected by each data sensor in the test platform have been pre-sampled to ensure the same number of sampling points. To construct a two-dimensional time-frequency feature map that meets the model input, the resampled one-dimensional time-domain signal is subjected to minimum-maximum normalization. This is then segmented into fixed-length signal segments and converted into a two-dimensional time-frequency feature map via a continuous wavelet transform. Finally, interpolation and resampling are performed to ensure that the signal meets the input requirements of the harmonic reducer fault diagnosis model.

[0042] In this implementation step, the expression of the signal to be analyzed is set to f(t), and the signal is normalized by minimum-maximum normalization to obtain the normalized signal f norm The formula for (t) is as follows:

[0043]

[0044] Among them, f max is the maximum value of the signal, f min is the minimum value of the signal.

[0045] The normalized multi-source heterogeneous data can be converted into a two-dimensional time-frequency domain feature map through continuous wavelet transform. Continuous wavelet transform belongs to the existing technology and is briefly described below.

[0046] Continuous wavelet transform obtains the spectrum information of the signal at different scales and positions by convolving the signal with a set of wavelet basis functions. The calculation formula of the wavelet transform is:

[0047]

[0048] The formula for continuous wavelet transform is defined as:

[0049]

[0050] Among them, a, b are the parameters of continuous transformation, a is the scale parameter, b is the translation parameter, f(t) is the original signal, which is a continuous time signal, ψ * is the complex conjugate of the wavelet function.

[0051] In the embodiment of the present invention, Morlet is selected as the wavelet basis function, and the expression of the Morlet wavelet basis function is:

[0052]

[0053] Where t is the time variable, ω0 is the frequency parameter that controls the frequency of the wavelet basis function, and j is the imaginary unit. The time variable is set to 256, a and b are set to 2 and 5 respectively, and the frequency parameter is set to 3.

[0054] The two-dimensional time-frequency domain feature map of each data sample in the above-mentioned multi-source heterogeneous data set after continuous wavelet transform can be used to construct training samples. Each training sample needs to include the two-dimensional time-frequency domain feature maps of three signals: vibration, torque, and temperature. It also needs to have the fault mode labels when these three signals are collected. Thus, all training samples can constitute a training sample set with fault mode labels. In an embodiment of the present invention, the training set and validation set can be divided with reference to the conventional model training method to facilitate obtaining the optimal model. In an embodiment of the present invention, in addition to dividing the training set and validation set, a portion of the training samples is also divided to verify the final performance of the model. Therefore, it can be divided into the training set, validation set, and test set of the network model in a ratio of 7:2:1. Of course, the specific division method can be adjusted with reference to the conventional model training method.

[0055] S3. Use the above-mentioned training sample set to train the harmonic reducer fault diagnosis model; the input of the harmonic reducer fault diagnosis model is the two-dimensional time-frequency domain feature map of three signals: vibration, torque and temperature. The two-dimensional time-frequency domain feature maps of the three signals are first subjected to noise suppression and feature selection through the deep residual shrinkage network, and then the important features are enhanced through the SE-Net module. Finally, the three enhanced features are fused at the feature level and the fault prediction label of the harmonic reducer is output through the prediction head.

[0056] In the harmonic reducer fault diagnosis model constructed by the present invention, it is necessary to input the preprocessed different two-dimensional time-frequency domain feature maps into the DRSN-SE model and extract features through the DRSN and SE-Net modules respectively. Finally, the features extracted by the three channels are fused at the feature level to ultimately realize multi-source heterogeneous data fusion under different fault modes.

[0057] like Figure 3As shown in the figure, the above-mentioned harmonic reducer fault diagnosis model contains three branches with the same structure. Each sensor signal corresponds to a branch. Each branch is composed of an input layer (Signal), two convolutional layers (Conv-2D), a maximum pooling layer (Maxpooling), three two-dimensional residual shrinkage unit (2D-RSUB) modules, a global average pooling layer (Global Average Pooling) and an SE-Net module cascaded in sequence. The input layer (Signal 1) of the first branch is used to input the two-dimensional time-frequency domain feature map of the vibration signal, the input layer (Signal 2) of the second branch is used to input the two-dimensional time-frequency domain feature map of the torque signal, and the input layer (Signal 3) of the third branch is used to input the two-dimensional time-frequency domain feature map of the harmonic temperature signal. After passing through the SE-Net module, each of the three branches obtains an enhanced feature; the enhanced features output by the three branches are fused by the feature fusion layer (Fusion) and then pass through the prediction head composed of a fully connected layer (Fully Connected) and a softmax layer to output the classification label of the fault mode.

[0058] The specific parameters of each network layer in the above network can be adjusted as needed. In an embodiment of the present invention, in each branch, the size of the two-time-frequency feature map input by the input layer is 3×256×256; the filter sizes of the two convolutional layers are 3 and 16 respectively, the convolution kernel size is 3×3, and the stride is 2; the convolution kernel size of the maximum pooling layer is 2×2, and the stride is 2; the size of the three filters in the first residual shrinkage unit (RSBU) is 32, the convolution kernel size is 3×3, and the stride is 1; the size of the three filters in the second RSBU is 48, the convolution kernel size is 3×3, and the stride is 1; the size of the three filters in the third RSBU is 64, the convolution kernel size is 3×3, and the stride is 1; the size of the feature map input to the global average pooling layer is 64×4×4, and the output size of the global average pooling layer is a tensor of 64×1×1.

[0059] The above DRSN contains multiple RSBUs, which integrate the soft threshold function and attention mechanism to achieve feature learning methods for strong noise or high redundancy data. The soft threshold function can adaptively remove noise and redundant information in the signal and enhance the learning ability of features. The expression of the soft threshold function is:

[0060]

[0061] Among them, x represents the input feature, y represents the output feature, and τ is the threshold.

[0062] Traditional RSBU is usually designed for one-dimensional time series signals. To adapt to the input samples, the one-dimensional convolution in RSBU is changed to a two-dimensional convolution. Figure 4As shown in the figure, in the two-dimensional residual shrinkage unit 2D-RSUB module, the input feature map of the module first passes through two two-dimensional convolutional layers (with BN operation and ReLu activation function) to obtain an intermediate feature map, and the intermediate feature map is subjected to absolute value operation (Absloute) and global average pooling (GAP) to obtain a C×1×1 dimensional feature vector Z; the feature vector Z passes through two fully connected layers (the second fully connected layer has a BN operation and ReLu activation function) to obtain a C×1×1 dimensional feature vector k, and the feature vector k is output through the Sigmoid function to obtain the scaling parameters of each channel (a total of C channels, each channel has a scaling parameter a c , whereby the Sigmoid function outputs a C×1×1 dimensional scaling parameter vector), all channels of the feature vector Z are scaled by their respective scaling parameters to obtain a threshold vector τ, and the threshold vector τ is used to perform soft threshold processing on the intermediate feature map to eliminate noise-related information in the input feature map; finally, the feature map after soft threshold processing is residually connected to the input feature map (identity shortcut) and passed to the next cascade network layer through the ReLU function.

[0063] The one-dimensional convolution in the above 2D-RSUB module has been changed to a two-dimensional convolution, but the main network layers are similar to the traditional RSBU, such as Absloute, scaling parameter calculation, and soft threshold processing. The specific calculation principle can be found in the traditional RSBU.

[0064] In the branch of the 2D-RSUB module, the core of noise reduction lies in soft threshold processing based on the scaling parameter. After the input features pass through two fully connected layers, the scaling parameter a is obtained through the output of the Sigmoid function, which is expressed as:

[0065]

[0066] Among them, k c represents the cth channel in the feature vector k output by the second FC layer, a c Represents the c-th channel of the scaling parameter a.

[0067] The calculation method of the threshold is expressed as:

[0068] τ c =a c ·Z

[0069] Among them, τ c is the threshold of the cth channel of the feature map, τ for all channels c Constitute the threshold vector τ; Z is the output feature of the convolution layer.

[0070] In this invention, the noise content of each sample input into the network model varies. The residual shrinkage unit automatically learns a set of thresholds based on the input samples. Each sample generates its own independent threshold based on its noise content to eliminate noise-related information. The final output is passed to the next layer of the network via the ReLU function, which accelerates the training process and improves computational efficiency.

[0071] The above SE-Net (Squeeze Excitation Network) module belongs to the existing technology and includes a convolution layer, a global average pooling layer, two fully connected layers, a Sigmoid activation function and a Scale unit. In the first convolution layer, the Transformation is performed and the F tr Convolution operation, constructing feature maps U, F tr The expression is

[0072]

[0073] X=[x 1 ,x 2 ,…,x C′ ]

[0074] in, * indicates convolution, is a 2D convolution kernel, V c The single channel that is applied will be applied to X.

[0075] In order to utilize channel dependencies, channel statistics are generated through global average pooling, and the W×H×C feature map containing global information is compressed into a 1×1×C feature vector Z.

[0076]

[0077] Among them, z c is the Cth element of Z.

[0078] The present invention uses aggregated information to fully capture channel dependencies and adopts a gating mechanism consisting of two fully connected layers. Through channel compression, nonlinear activation function and channel number recovery, the sigmoid activation function is used to obtain a weight s with a dimension of 1×1×C to characterize the weight of the channel in the feature map U. The expression of s is:

[0079] s=F ex (z,W)=σ(g(z,W))=σ(W2δ(W1z))

[0080] Where δ refers to the ReLU function, r represents the scaling factor.

[0081] Finally, the attention weights obtained above are weighted to the features of each channel through the Scale operation to obtain the final output of the SE-Net module.

[0082]

[0083] in, F scale (u c ,s c ) refers to the feature map With scalar s c Channel multiplication.

[0084] In the global average pooling layer, the size of the input feature map is 64×1×1. The features of each channel are averaged to generate a 1×1×64 feature map. The number of channels C input to the first fully connected layer is 64. Assuming the scaling factor is r=16, the output channels after the first fully connected layer are 4. After the ReLU function and the second fully connected layer, the number of channels output at the Scale layer is 64.

[0085] After the data of the three modalities pass through the same network structure, feature splicing operation is performed in the feature fusion layer to splice and integrate the features of the three branches, and feature classification and recognition are realized through the Softmax layer.

[0086] The DRSN-SE model was developed using the Python programming language within the Pytorch deep learning framework. During training, the learning rate was set to 0.001, and the Adam algorithm was used for optimization. The batch size was set to 90, and the number of epochs was set to 40. The parameters of each neural network layer are shown in Table 1, where the first number in the brackets in the first column indicates the number of channels, the middle number indicates the size of the convolution kernel, and the last number indicates the step size. The second column represents the number of units, and the third column indicates the output size of each layer, in the format of number of channels × width × height.

[0087] Table 1 DRSN-SE model parameter settings

[0088]

[0089]

[0090] The constructed dataset is used to train and verify the DRSN-SE model, the trained model is tested and the optimal model parameters are saved.

[0091] S4. For harmonic reducers with unknown fault modes, one-dimensional time domain signals of vibration, torque and temperature can be obtained and converted into two-dimensional time-frequency domain feature maps respectively through the continuous wavelet transform method, and then input into the trained harmonic reducer fault diagnosis model to predict the fault mode of the harmonic reducer.

[0092] In this embodiment of the present invention, the test set data was used to verify the performance of the DRSN-SE model, demonstrating that the model can effectively diagnose faults in harmonic reducers. The trained model parameters were then saved. This model can be integrated into the data processing equipment of the robot joint integration module, allowing the model to integrate fault warning functions into other devices.

[0093] The embodiment described above is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Persons skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent substitution or equivalent transformation falls within the scope of protection of the present invention.

Claims

1. A harmonic reducer fault diagnosis method based on multi-source heterogeneous data fusion, characterized in that: The steps include: S1. Data sensors placed on the harmonic reducer are used to collect one-dimensional time domain signals of vibration, torque, and temperature of the harmonic reducer under different fault modes. The three one-dimensional time domain signals collected under each fault mode constitute a data sample in the multi-source heterogeneous dataset. S2. Preprocess each data sample in the multi-source heterogeneous data, convert the one-dimensional time domain signals of vibration, torque, and temperature into two-dimensional time-frequency domain feature maps using the continuous wavelet transform method, thereby converting each data sample into a training sample and constructing a training sample set with fault mode labels; S3. Using the training sample set to train a harmonic reducer fault diagnosis model; the input of the harmonic reducer fault diagnosis model is a two-dimensional time-frequency domain feature map of three signals: vibration, torque, and temperature. The two-dimensional time-frequency domain feature maps of the three signals are first subjected to noise suppression and feature selection by a deep residual shrinkage network, and then the important features are enhanced by the SE-Net module. Finally, the three enhanced features are fused at the feature level and a prediction head is used to output a fault prediction label for the harmonic reducer; S4. For a harmonic reducer with an unknown fault mode, obtain the one-dimensional time domain signals of vibration, torque and temperature and convert them into two-dimensional time-frequency domain feature maps respectively through the continuous wavelet transform method, and then input them into the trained harmonic reducer fault diagnosis model to predict the fault mode of the harmonic reducer.

2. The harmonic reducer fault diagnosis method based on multi-source heterogeneous data fusion according to claim 1 is characterized in that: The data sensors arranged on the harmonic reducer include an acceleration sensor for detecting vibration signals, a torque sensor for detecting the output torque of the harmonic reducer, and a temperature sensor for detecting the operating temperature of the harmonic reducer.

3. The harmonic reducer fault diagnosis method based on multi-source heterogeneous data fusion according to claim 2 is characterized in that: The acceleration sensor is adsorbed on the outer surface of the rigid wheel by magnetic attraction; the torque sensor is installed at the output end of the harmonic reducer; the temperature sensor uses a temperature probe and is attached to the surface of the harmonic reducer to measure the real-time operating temperature.

4. The harmonic reducer fault diagnosis method based on multi-source heterogeneous data fusion according to claim 1 is characterized in that: The failure modes covered by the data samples in the multi-source heterogeneous data set include six types: no fault, rigid wheel wear fault, flexspline wear fault, flexspline pitting fault, flexspline crack fault, and bearing pitting fault.

5. The harmonic reducer fault diagnosis method based on multi-source heterogeneous data fusion according to claim 1 is characterized in that: For the one-dimensional time domain signal collected by each data sensor, it needs to be resampled and minimum-maximum normalized, and then divided into signal segments of fixed length and converted into a two-dimensional time-frequency feature map through continuous wavelet transform. The two-dimensional time-frequency feature map obtained by continuous wavelet transform is resampled by interpolation to make it meet the input requirements of the harmonic reducer fault diagnosis model.

6. The harmonic reducer fault diagnosis method based on multi-source heterogeneous data fusion according to claim 1 is characterized in that: The harmonic reducer fault diagnosis model includes three branches with the same structure, each sensor signal corresponds to a branch, and each branch is composed of an input layer, two convolution layers, a maximum pooling layer, three two-dimensional residual shrinkage unit modules, a global average pooling layer and an SE-Net module cascaded in sequence. The input layer of the first branch is used to input the two-dimensional time-frequency domain feature map of the vibration signal, the input layer of the second branch is used to input the two-dimensional time-frequency domain feature map of the torque signal, and the input layer of the third branch is used to input the two-dimensional time-frequency domain feature map of the harmonic temperature signal. After passing through the SE-Net module, each of the three branches obtains an enhanced feature; The enhanced features output by the three branches are fused through the feature fusion layer and then pass through the prediction head composed of a fully connected layer and a Softmax layer to output the classification label of the fault mode.

7. The harmonic reducer fault diagnosis method based on multi-source heterogeneous data fusion according to claim 6 is characterized in that: In each of the two-dimensional residual shrinkage unit modules, the module's input feature map first passes through two two-dimensional convolutional layers to obtain an intermediate feature map, and the intermediate feature map is subjected to absolute value operation and global average pooling to obtain a feature vector; the feature vector passes through two fully connected layers and is output through a sigmoid function to obtain a scaling parameter for each channel; all channels of the feature vector are scaled by their respective scaling parameters to obtain a threshold vector, and the threshold vector is used to perform soft threshold processing on the intermediate feature map to eliminate noise-related information in the input feature map; Finally, the feature map after soft threshold processing is connected to the residual of the input feature map and passed to the next cascade network layer through the ReLU function.

8. The method for fault diagnosis of harmonic reducer based on multi-source heterogeneous data fusion according to claim 6, characterized in that: In each branch of the harmonic reducer fault diagnosis model, the size of the two-time-frequency feature map input by the input layer is 3×256×256; the filter sizes of the two convolutional layers are 3 and 16 respectively, the convolution kernel size is 3×3, and the stride is 2; the convolution kernel size of the maximum pooling layer is 2×2, and the stride is 2; the size of the three filters in the first residual contraction unit is 32, the convolution kernel size is 3×3, and the stride is 1; the size of the three filters in the second residual contraction unit is 48, the convolution kernel size is 3×3, and the stride is 1; the size of the three filters in the third residual contraction unit is 64, the convolution kernel size is 3×3, and the stride is 1; the feature map size of the input global average pooling layer is 64×4×4, and the output size of the global average pooling layer is a tensor of 64×1×1.

9. The harmonic reducer fault diagnosis method based on multi-source heterogeneous data fusion according to claim 6, characterized in that: In the feature fusion layer, the three enhanced features are fused at the feature level through a feature splicing operation.

10. The harmonic reducer fault diagnosis method based on multi-source heterogeneous data fusion according to claim 1, characterized in that: When the harmonic reducer fault diagnosis model is trained using the training sample set, it is necessary to divide it into a training set and a validation set in advance according to a preset ratio.

Images