A mechanical equipment fault diagnosis method based on multi-signal fusion

Abstract

This invention discloses a mechanical equipment fault diagnosis method based on multi-signal fusion, belonging to the field of equipment fault diagnosis method design, and includes the following steps: establishing an operating signal set. U and running status set V Based on the characteristics of the operating signals, the sampling frequency is determined, with a larger sampling frequency used for time-sensitive operating signals and a smaller sampling frequency used for those not sensitive to time changes. Noise reduction processing is performed on the operating signals. The operating signals are then normalized to decimal values ​​between 0 and 1. A sample set is constructed based on the processed operating signals. A fault diagnosis model is built, and its parameters are set. The fault diagnosis model is trained. Fault diagnosis is then performed, and the results are output. Compared to traditional methods of judging mechanical equipment faults using a single signal, this method uses multiple signals for fault diagnosis, considering both historical data and the current situation, resulting in more accurate results.

Description

Technical Field

[0001] This invention relates to the field of equipment fault diagnosis technology, and specifically to a method for diagnosing mechanical equipment faults based on multi-signal fusion. Background Technology

[0002] Fault diagnosis technology refers to the techniques used to determine whether mechanical equipment is operating normally and to promptly detect faults based on its operational status. It is a powerful guarantee for the safe production and efficient operation of mechanical equipment. In automated production, more accurate and intelligent fault diagnosis of mechanical equipment is of great significance for improving industrial production efficiency and economic benefits. With the rapid development of industry, the structures of various production equipment are becoming increasingly complex, and their functions are becoming increasingly diversified, inevitably leading to an increase in equipment failure rates. Because the failures of these complex equipment are often multi-source, complex, and concealed, finding the causes of failures using traditional fault diagnosis methods is becoming increasingly difficult. When mechanical equipment malfunctions, it will exhibit various symptoms, such as changes in vibration, temperature, and noise signals. Existing fault diagnosis methods often make judgments based on a single vibration signal, without considering the influence of other signals on the diagnostic results, resulting in a lack of accuracy.

[0003] Patent document CN109506921B discloses an online diagnosis and fault early warning method for rotating machinery. After collecting vibration data, the method immediately analyzes and processes the data to obtain the characteristic parameters of each component during the operation of the rotating machinery. However, it only uses the vibration data of the machinery to diagnose the faults of the mechanical equipment and does not utilize a multi-faceted set of diagnostic signals, so the diagnostic results are still not accurate enough. Summary of the Invention

[0004] The purpose of this invention is to provide a mechanical equipment fault diagnosis method based on multi-signal fusion, which makes the diagnosis results more scientific and reasonable, so as to solve the shortcomings of the existing technology that makes judgments based on a single vibration signal with low accuracy.

[0005] To achieve the above objectives, the present invention employs the following technical solution: a method for fault diagnosis of mechanical equipment based on multi-signal fusion, comprising the following steps:

[0006] S1: Establish the operating signal set U and the operating state set V;

[0007] S2: Determine the sampling frequency based on the characteristics of the operating signal. Use a larger sampling frequency for operating signals that are sensitive to time changes, and a smaller sampling frequency for operating signals that are not sensitive to time changes.

[0008] S3: Perform noise reduction processing on the operating signal after S2 processing;

[0009] S4: Normalize the running signal after processing by S3, so that its value becomes a decimal between 0 and 1.

[0010] S5: Process the running signal after S4 to construct a sample set;

[0011] S6: Construct a fault diagnosis model and set the fault diagnosis model parameters;

[0012] S7: Train the fault diagnosis model;

[0013] S8: Fault diagnosis, outputs the results of fault diagnosis.

[0014] The operating signal set U and operating status set V in S1 are established based on the structural characteristics of the mechanical equipment being diagnosed.

[0015] Running signal set U = {u1, u2, ..., u} m}, where u1, u2, ..., u m This represents different operating signals of the device, where m is the number of operating signals; establish the operating state set V = {v1, v2, ..., v...} n}, v1, v2, ..., v n This represents the running state, where n is the number of running states.

[0016] The noise reduction process in S3 specifically includes:

[0017] S3-1: The acquired running signal is decomposed using Complete Ensemble Empirical Mode Decomposition With Adaptive Noise (CEEMDAN) to obtain n intrinsic mode functions and one residual signal. The n intrinsic mode functions are represented as IMF1, IMF2, ..., IMF1. n The residual signal is represented as r;

[0018] S3-2: The correlation coefficients between the n intrinsic mode functions and the original signal are calculated using the Pearson correlation coefficient. The calculation method for the Pearson correlation coefficient is shown in formula (1).

[0019]

[0020] In the formula: x represents the original signal, IMF i Let i represent the IMF component of level i. According to formula (1), when the original signal and the IMF component are negatively correlated, the correlation coefficient is negative, and the higher the linear correlation, the larger the correlation coefficient.

[0021] S3-3: Signal reconstruction to obtain the denoised signal;

[0022] When the correlation coefficient between the intrinsic mode function and the original signal When the value is less than a set threshold, the intrinsic mode function is identified as noise and removed.

[0023] When the correlation coefficient between the intrinsic mode function and the original signal When the intrinsic mode function is greater than or equal to a set threshold, it is identified as the signal generated and is retained. The retained intrinsic mode function is then reconstructed with the residual signal to form a denoised signal.

[0024] The specific construction method of the sample set in S5 is as follows:

[0025] S5-1: Operating signals that are sensitive to time changes have a higher sampling frequency and more sampled data. Starting from the acquisition start point, several consecutive data points are taken as a sample. Operating signals that are not sensitive to time changes have a lower sampling frequency and less sampled data. Each data point is taken as a sample.

[0026] S5-2: Process the time-sensitive operating signal samples to obtain a two-dimensional time-frequency diagram. Use CEEMDAN to decompose the time-sensitive operating signal, and then use Hilbert transform (HT) to perform Hilbert transform on each IMF component after CEEMDAN decomposition to generate a two-dimensional time-frequency diagram.

[0027] The fault diagnosis model in S6 includes: a 2D-CNN channel, a 1D-CNN channel, and a fusion channel;

[0028] The 2D-CNN channel includes an input layer (input2d), a two-dimensional convolutional layer (conv2d), a two-dimensional max pooling layer (max_pooling2d), a two-dimensional convolutional layer (conv2d_1), a two-dimensional max pooling layer (max_pooling2d_1), an attention mechanism module layer (se_attention_layer), a two-dimensional convolutional layer (conv2d_2), a two-dimensional max pooling layer (max_pooling2d_2), and a flatten layer.

[0029] 1D-CNN includes an input layer (input1d), a one-dimensional convolutional layer (conv1d), a one-dimensional max pooling layer (max_pooling1d), a one-dimensional convolutional layer (conv1d_1), a one-dimensional max pooling layer (max_pooling1d_1), an attention layer (attention_layer), a one-dimensional convolutional layer (conv1d_2), a one-dimensional max pooling layer (max_pooling1d_2), and a flatten layer (flatten_1).

[0030] The fusion channel includes a concatenate layer, a dropout layer, a feature layer, and an output layer.

[0031] The time-sensitive running signal is processed to obtain a two-dimensional time-frequency graph, which is then stitched together and used as the input to the 2D-CNN channel input layer input2d; the running signal is stitched together and used as the input to the 1D-CNN channel input layer input1d; the 2D-CNN channel flatten layer flatten and the 1D-CNN channel flatten_1 layer flatten_1 are used as the input to the concatenate layer.

[0032] The diagnostic model for S7 is established by the following steps:

[0033] S7-1: Divide the sample set into a training set and a test set according to a certain ratio, with the ratio of the number of samples in the training set to the number of samples in the test set being 7:3;

[0034] S7-2: Set training parameters. The Adam optimizer was used to update the network parameters. The Epoch, batch size, and learning rate of the Adam algorithm during model training were 60, 64, and 0.001, respectively.

[0035] S7-3: Model training. The fault diagnosis model parameters are initialized using a random method. Training samples are input into the fault diagnosis model in batches for training. Test samples are input into the trained fault diagnosis model for testing. The fault diagnosis accuracy is output to determine the feasibility of the proposed model. The trained fault diagnosis model is saved after training.

[0036] In S8, the fault diagnosis specifically involves inputting the set of operating signals U to be diagnosed into the trained fault diagnosis model, performing the diagnosis, and outputting the diagnosis results.

[0037] According to the above technical solution, the present invention has at least the following beneficial effects:

[0038] 1. In the process of diagnosing mechanical equipment faults, this invention considers the influence of three signals—temperature, vibration, and noise—on the accuracy of fault diagnosis, thus avoiding the shortcomings of current mechanical equipment fault diagnosis methods that rely on a single vibration signal, resulting in low accuracy.

[0039] 2. In collecting vibration, temperature and noise signals, this invention considers the sensitivity of the three variables to time changes and uses different sampling frequencies, which ensures the integrity of the collected data and reduces the amount of computation.

[0040] 3. The dual-channel convolutional neural network fault diagnosis model disclosed in this invention learns the correlation between adjacent intervals and the correlation between non-adjacent intervals, thereby improving the accuracy of fault diagnosis and avoiding the shortcomings of current mechanical equipment fault diagnosis models that use a single channel, resulting in low fault diagnosis accuracy.

[0041] 4. When performing fault diagnosis, this invention considers both historical data and the current situation, resulting in more accurate results compared to traditional fault diagnosis methods based on the current situation.

[0042] 5. The dual-channel convolutional neural network fault diagnosis model disclosed in this invention adopts a combination of CEEMDAN decomposition and HT method in the time-frequency transformation module of its two-dimensional convolutional channel, avoiding the problems of mode aliasing caused by the original signal decomposition by EMI in traditional HHT transformation, or the difficulty in determining a suitable wavelet basis in wavelet transform. Attached Figure Description

[0043] Figure 1 This is a flowchart of the fault diagnosis steps in this example;

[0044] Figure 2 This is a diagram of the dual-channel fault diagnosis model for this example;

[0045] Figure 3 It is a two-dimensional time-frequency diagram of a vibration signal sample under normal operating conditions;

[0046] Figure 4 It is a two-dimensional time-frequency diagram of a vibration signal sample under end-face wear conditions;

[0047] Figure 5 It is a two-dimensional time-frequency diagram of a vibration signal sample under the condition of slipper wear;

[0048] Figure 6 This is a two-dimensional time-frequency diagram of a vibration signal sample under swashplate wear conditions.

[0049] Figure 7 This is a two-dimensional time-frequency diagram of a vibration signal sample under the condition of wear on the distribution plate.

[0050] Figure 8 This is a two-dimensional time-frequency diagram of a vibration signal sample under the condition of central spring failure.

[0051] Figure 9 This is a graph showing the changes in the loss function value and accuracy of the training set samples in this invention.

[0052] Figure 10 This is a comparison chart of the accuracy rates of different fault diagnosis methods. Detailed Implementation

[0053] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0054] The flowchart of the mechanical equipment fault diagnosis method based on multi-signal fusion proposed in this invention is as follows. Figure 1 As shown below, the specific implementation steps are illustrated using a hydraulic pump fault diagnosis as an example:

[0055] S1. Establish the running signal set U and the running status set V.

[0056] Based on the structural characteristics of the mechanical equipment being diagnosed, an operating signal set is established, U = {u1, u2, ..., u...}. m}, where u1, u2, ..., u m This represents different operating signals of the device, where m is the number of operating signals; the operating state set V = {v1, v2, ..., v...} n}, v1, v2, ..., v n This represents the running state, where n is the number of running states.

[0057] For this hydraulic pump, the operating signals include temperature signals, vibration signals, and noise signals. That is, the operating signal set U = {u1, u2, ..., u...} m} = {temperature signal, vibration signal, noise signal}.

[0058] During the operation of this hydraulic pump, possible abnormal states include end face wear, slipper wear, swashplate wear, distributor plate wear, and center spring failure, as well as normal states. Therefore, the operating state set V = {v1, v2, ..., v...} n} = {Normal state, end face wear, slipper wear, swashplate wear, distributor plate wear, center spring failure}.

[0059] S2. Determine the sampling frequency based on the characteristics of the operating signal. Use a larger sampling frequency for operating signals that are sensitive to time changes, and a smaller sampling frequency for operating signals that are not sensitive to time changes.

[0060] For the fault diagnosis of the hydraulic pump in this embodiment, the vibration signal and noise signal are sensitive to time changes, and their sampling frequency is 0.1K Hz. The temperature signal is not sensitive to time changes, and its sampling frequency is 0.1Hz.

[0061] S3. Noise reduction processing of operating signals

[0062] S3-1 uses Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to decompose the acquired running signal, obtaining n Intrinsic Mode Functions (IMFs) and a residual signal. The n IMFs are denoted as IMF1, IMF2, ..., IMF1. n The residual signal is represented as r;

[0063] S3-2 uses the Pearson correlation coefficient to calculate the correlation coefficients between the n intrinsic mode functions (IMFs) and the original signal. The Pearson correlation coefficient calculation method is shown in formula (1);

[0064]

[0065] In the formula: x represents the original signal, IMF i Let i represent the IMF component of level i. According to formula (1), when the original signal and the IMF component are negatively correlated, the correlation coefficient is negative, and the higher the linear correlation, the larger the correlation coefficient.

[0066] S3-3 signal reconstruction yields the denoised signal. The correlation coefficient between the intrinsic mode function and the original signal is then calculated. When the intrinsic mode function (IMF) is less than a set threshold, it is considered to be generated by noise and is removed; when the correlation coefficient between the IMF and the original signal is less than a set threshold, it is removed. When the intrinsic mode function is greater than or equal to a set threshold, it is identified as the signal generated and is retained. The retained intrinsic mode function is then reconstructed with the residual signal to form a denoised signal.

[0067] For the fault diagnosis of the hydraulic pump in this embodiment, the vibration signal and noise signal are decomposed by CEEMDAN to obtain their respective intrinsic mode functions. Their correlation coefficients are calculated, and one-tenth of the maximum correlation coefficient in the correlation coefficient sequence is used as the threshold to remove the intrinsic mode functions generated by noise. After reconstruction, the denoised vibration signal and noise signal are obtained.

[0068] S4. Normalize the running signal to make its value a decimal between (0, 1).

[0069] S5. Process the running signals and construct a sample set;

[0070] S5-1 samples time-sensitive operating signals at a higher frequency, resulting in a larger amount of data. Starting from the initial acquisition point, several consecutive data points are collected as a single sample. For operating signals that are not time-sensitive, the sampling frequency is lower, resulting in a smaller amount of data. Each data point is collected as a single sample.

[0071] For the fault diagnosis of the hydraulic pump in this embodiment, for each state, 1024 data points of vibration signal and noise signal are taken as a sample starting from the starting point, then another 1024 data points are taken as the next sample, and so on, with 600 samples taken for each state. There are 6 states for the hydraulic pump, so a total of 3600 samples of vibration signal and 3600 samples of noise signal are taken. For each state, the temperature signal is taken as a sample starting from the starting point, with 600 samples taken for each state. There are 6 states for the hydraulic pump, so a total of 3600 samples of temperature signal are taken.

[0072] S5-2 processes time-sensitive operating signal samples to obtain a two-dimensional time-frequency graph. The time-sensitive operating signal is decomposed using CEEMDAN, and then Hilbert transform (HT) is applied to each IMF component after CEEMDAN decomposition to generate a two-dimensional time-frequency graph.

[0073] In this example, the vibration signal sample and noise signal sample are first decomposed by CEEMDAN, and then the Hilbert transform (HT) is used to perform Hilbert transform on each IMF component after CEEMDAN decomposition to generate a two-dimensional time-frequency diagram.

[0074] S6. Construct a fault diagnosis model and set the fault diagnosis model parameters;

[0075] like Figure 2 As shown, the fault diagnosis model includes 2D-CNN channels, 1D-CNN channels, and fusion channels.

[0076] The 2D-CNN channel includes an input layer (input2d), a 2D convolutional layer (conv2d), a 2D max pooling layer (max_pooling2d), a 2D convolutional layer (conv2d_1), a 2D max pooling layer (max_pooling2d_1), an attention mechanism module layer (se_attention_layer), a 2D convolutional layer (conv2d_2), a 2D max pooling layer (max_pooling2d_2), and a flatten layer.

[0077] 1D-CNN includes an input layer (input1d), a one-dimensional convolutional layer (convld), a one-dimensional max pooling layer (max_pooling1d), a one-dimensional convolutional layer (conv1d_1), a one-dimensional max pooling layer (max_poolingld_1), an attention layer (attention_layer), a one-dimensional convolutional layer (conv1d_2), a one-dimensional max pooling layer (max_pooling1d_2), and a flatten layer (flatten_1).

[0078] The fusion channel includes the concatenation layer, the dropout layer, the feature layer, and the output layer.

[0079] The time-sensitive running signal is processed to obtain a two-dimensional time-frequency graph, which is then stitched together and used as the input to the 2D-CNN channel input layer input2d; the running signal is stitched together and used as the input to the 1D-CNN channel input layer input1d; the 2D-CNN channel flatten layer flatten and the 1D-CNN channel flatten_1 layer flatten_1 are used as the input to the concatenate layer.

[0080] In this embodiment, the input layer `input2d` has a size of 32×32×6; the 2D convolutional layer `conv2d` has a kernel size of 3×3, 4 kernels, ReLU activation function, stride of 1, and same padding method; the 2D max pooling layer `max_pooling2d` has a kernel size of 2×2 and stride of 2; the 2D convolutional layer `conv2d_1` has a kernel size of 5×5, 8 kernels, ReLU activation function, stride of 1, and same padding method; the 2D max pooling layer `max_pooling2d_1` has a kernel size of 2×2 and stride of 2; the attention mechanism module layer `se_attention_layer` uses SENet; the 2D convolutional layer `conv2d_2` has a kernel size of 3×3, 16 kernels, ReLU activation function, stride of 1, and same padding method; the 2D max pooling layer `max_pooling2d_2` has a kernel size of 2×2 and stride of 2.

[0081] In this embodiment, the input layer `input1d` has a size of 2049×1; the one-dimensional convolutional layer `conv1d` has a kernel size of 7×1, two kernels, a ReLU activation function, a stride of 1, and the same padding method; the one-dimensional max pooling layer `max_pooling1d` has a kernel size of 3×1 and a stride of 3; the one-dimensional convolutional layer `conv1d_1` has a kernel size of 5×1, four kernels, a ReLU activation function, a stride of 1, and the same padding method. The one-dimensional max pooling layer (max_pooling1d_1) has a kernel size of 2×1 and a stride of 2. The attention mechanism module layer (attention_layer) uses BahdanauAttention. The one-dimensional convolutional layer (conv1d_2) has a kernel size of 3×1, 8 kernels, ReLU activation function, stride of 1, and same padding method. The one-dimensional max pooling layer (max_pooling1d_2) has a kernel size of 4×1 and a stride of 4.

[0082] In this embodiment, the number of neurons in the splicing layer is 936, the dropout rate of the dropout layer is 0.5, the number of neurons in the dense layer is 128, the activation function is ReLU, the number of neurons in the output layer is 6, and the activation function is Softmax.

[0083] In this example, the feature map formed by stitching together the two-dimensional time-frequency graphs of vibration signal and noise signal is used as the input of the 2D-CNN channel input layer input2d, and the data vector formed by stitching together temperature signal, vibration signal sample and noise signal is used as the input of the 1D-CNN channel input layer input1d.

[0084] S7. Train the fault diagnosis model;

[0085] S7-1 divides the data samples into training and testing sets according to a certain ratio;

[0086] In this example, there are 600 samples for each state, and a total of 6 states. The 3600 feature maps formed by stitching together the two-dimensional time-frequency graphs of vibration and noise signals are divided into 3600 data vectors after stitching together temperature, vibration, and noise signals in a 7:3 ratio. The number of training set samples is 2520, and the number of test set samples is 180.

[0087] S7-2 sets training parameters;

[0088] In this example, the Adam optimizer is used to update the network parameters. The epoch, batch size, and learning rate of the Adam algorithm during model training are 60, 64, and 0.001, respectively.

[0089] S7-3 model training;

[0090] The fault diagnosis model parameters are initialized using a random method. Training samples are input into the fault diagnosis model in batches for training, and test samples are input into the trained fault diagnosis model for testing. The fault diagnosis accuracy is output to determine the feasibility of the proposed model. The trained fault diagnosis model is saved after training is complete.

[0091] S8, Fault Diagnosis, outputs the results of fault diagnosis;

[0092] The temperature, vibration, and noise signals that need to be diagnosed are input into the trained fault diagnosis model to perform the diagnosis and output the diagnosis results.

[0093] The effectiveness of this invention can be further illustrated by the following comparative experiments:

[0094] 1) The hydraulic pump was run for a period of time under normal conditions, end face wear, slipper wear, swashplate wear, distributor plate wear, and center spring failure conditions. The temperature signal was sampled by a temperature sensor at a sampling frequency of 0.1 Hz, the vibration signal was sampled by an acceleration sensor at a sampling frequency of 0.1 kHz, and the noise signal was sampled by a noise sensor at a sampling frequency of 0.1 kHz.

[0095] 2) The collected vibration and noise signals are denoised. The collected vibration and noise signals are first decomposed by CEEEMDAN, and then the Pearson correlation coefficient between each decomposed IMF component and its corresponding original signal is calculated using formula (1). One-tenth of the maximum correlation coefficient in the correlation coefficient sequence is used as the threshold to remove IMF components with Pearson correlation coefficients less than the threshold, and the remaining components are reconstructed.

[0096] 3) Normalize the reconstructed decomposition to form normalized temperature signal, vibration signal, and noise signal.

[0097] 4) 3600 temperature signal samples, 3600 vibration signal samples, and 3600 noise signal samples were collected.

[0098] For the normalized temperature signals continuously collected by the hydraulic pump under normal operating conditions, starting from the starting point, each data point is taken as a sample, for a total of 600 temperature signal samples under normal operating conditions. For the normalized temperature signals continuously collected by the hydraulic pump under end-face wear conditions, starting from the starting point, each data point is taken as a sample, for a total of 600 temperature signal samples under end-face wear conditions. For the normalized temperature signals continuously collected by the hydraulic pump under slipper wear conditions, starting from the starting point, each data point is taken as a sample, for a total of 600 temperature signal samples under slipper wear conditions. For the normalized temperature signals continuously collected by the hydraulic pump under swashplate wear conditions, starting from the starting point, each data point is taken as a sample, for a total of 600 temperature signal samples under swashplate wear conditions. For the normalized temperature signals continuously collected by the hydraulic pump under the condition of distributor plate wear, starting from the starting point, each data point is taken as a sample, for a total of 600 temperature signal samples under the condition of center spring failure.

[0099] For the normalized vibration signals continuously collected from the hydraulic pump under normal operating conditions, 1024 data points are taken as one sample starting from the starting point, then another 1024 data points are taken as the next sample, and so on, for a total of 600 vibration signal samples under normal operating conditions. For the normalized vibration signals continuously collected from the hydraulic pump under end-face wear conditions, 1024 data points are taken as one sample starting from the starting point, then another 1024 data points are taken as the next sample, and so on, for a total of 600 vibration signal samples under end-face wear conditions. For the normalized vibration signals continuously collected from the hydraulic pump under slipper wear conditions, 1024 data points are taken as one sample starting from the starting point, then another 1024 data points are taken as the next sample, and so on, for a total of 600 vibration signal samples under slipper wear conditions. For the normalized vibration signals of the hydraulic pump continuously collected under swashplate wear conditions, 1024 data points are taken as a sample starting from the starting point, then another 1024 data points are taken as the next sample, and so on, for a total of 600 vibration signal samples under swashplate wear conditions. For the normalized vibration signals of the hydraulic pump continuously collected under distributor plate wear conditions, 1024 data points are taken as a sample starting from the starting point, then another 1024 data points are taken as the next sample, and so on, for a total of 600 vibration signal samples under center spring failure conditions. For the normalized vibration signals of the hydraulic pump continuously collected under distributor plate wear conditions, 1024 data points are taken as a sample starting from the starting point, then another 1024 data points are taken as the next sample, and so on, for a total of 600 vibration signal samples under center spring failure conditions.

[0100] For the normalized noise signal continuously collected from the hydraulic pump under normal operating conditions, 1024 data points are taken as one sample starting from the starting point, then another 1024 data points are taken as the next sample, and so on, for a total of 600 noise signal samples under normal operating conditions. For the normalized noise signal continuously collected from the hydraulic pump under end-face wear conditions, 1024 data points are taken as one sample starting from the starting point, then another 1024 data points are taken as the next sample, and so on, for a total of 600 noise signal samples under end-face wear conditions. For the normalized noise signal continuously collected from the hydraulic pump under slipper wear conditions, 1024 data points are taken as one sample starting from the starting point, then another 1024 data points are taken as the next sample, and so on, for a total of 600 noise signal samples under slipper wear conditions. For the normalized noise signals continuously collected from the hydraulic pump operating under swashplate wear conditions, 1024 data points are taken as one sample starting from the starting point, then another 1024 data points are taken as the next sample, and so on, for a total of 600 noise signal samples under swashplate wear conditions. For the normalized noise signals continuously collected from the hydraulic pump operating under distributor plate wear conditions, 1024 data points are taken as one sample starting from the starting point, then another 1024 data points are taken as the next sample, and so on, for a total of 600 noise signal samples under center spring failure conditions. For the normalized noise signals continuously collected from the hydraulic pump operating under distributor plate wear conditions, 1024 data points are taken as one sample starting from the starting point, then another 1024 data points are taken as the next sample, and so on, for a total of 600 noise signal samples under center spring failure conditions.

[0101] 5) The 3600 temperature signal samples and 3600 vibration signal samples were first decomposed using CEEMDAN, and then Hilbert transform (HT) was applied to each IMF component after CEEMDAN decomposition to generate a two-dimensional time-frequency plot. Two-dimensional time-frequency plots of some vibration signal samples are shown in Figures 3 to 34. Figure 8 As shown.

[0102] 6) The two-dimensional time-frequency images of 3600 temperature signal samples are stitched together with the corresponding two-dimensional time-frequency images of 3600 vibration signal samples. After the images are compressed using a bicubic interpolation algorithm, 3600 32×32×6 input feature maps are formed. The corresponding 3600 temperature signal samples, 3600 vibration signal samples, and 3600 noise signal samples are stitched together to form 3600 2049×1 data vectors.

[0103] 7) 3600 32×32×6 input feature maps and 3600 2049×1 data vectors are randomly divided into training and test sets in a 7:3 ratio.

[0104] 8) Input the training set samples into the fault diagnosis model, such as... Figure 9 As shown, the model was trained. After training, the training effect was tested using a test set. The training effect is shown in Table 1. State 1, State 2, State 3, State 4, State 5, and State 6 represent the normal state, end face wear, slipper wear, swashplate wear, distributor plate wear, and center spring failure state, respectively. The accuracy of the diagnostic results is 99.44%.

[0105] Table 1 Test Results of Test Set Samples

[0106]

[0107] like Figure 10 As shown, M1, M2, M3, M4, M5, M6, M7, M8, and M9 represent the fault diagnosis methods disclosed in this invention: a multi-information fusion 2D-CNN single-channel fault diagnosis method, a multi-information fusion 1D-CNN single-channel fault diagnosis method, a vibration signal dual-channel fault diagnosis method, a noise signal dual-channel fault diagnosis method, a vibration signal 2D-CNN single-channel fault diagnosis method, a vibration signal 1D-CNN single-channel fault diagnosis method, a noise signal 2D-CNN single-channel fault diagnosis method, and a noise signal 1D-CNN single-channel fault diagnosis method. To demonstrate the superiority of the dual-channel fault diagnosis method compared to the single-channel fault diagnosis model, experiments were conducted on the 2D-CNN channel model and the 1D-CNN channel model using the same dataset, achieving diagnostic accuracies of 94.44% and 95.56%, respectively. To demonstrate the superiority of the fault diagnosis method disclosed in this invention, which utilizes multi-information fusion compared to fault diagnosis models based on a single vibration signal or a single noise signal, experimental tests were conducted on the fault diagnosis model using single vibration signals and single noise signals as sample sets. The diagnostic accuracies were 93.33% and 92.22%, respectively. To further demonstrate the overall superiority of the fault diagnosis model, experimental tests were conducted on a 2D-CNN channel model and a 1D-CNN channel model using single vibration signals and single noise signals, respectively. The diagnostic accuracies were 88.33%, 87.22%, 84.44%, and 83.89%, respectively.

Claims

1. A mechanical equipment fault diagnosis method based on multi-signal fusion, characterized in that, Includes the following steps: S1: Establish the operating signal set U and the operating state set V; S2: Determine the sampling frequency based on whether the operating signal is sensitive to time changes; S3: Perform noise reduction processing on the operating signal after S2 processing; S4: Normalize the running signal after processing by S3, so that its value becomes a decimal between 0 and 1. S5: Process the running signal after S4 to construct a sample set; S6: Construct a fault diagnosis model and set the fault diagnosis model parameters; S7: Train the fault diagnosis model; S8: Fault diagnosis, outputs the results of fault diagnosis; The operating signal set U and operating status set V in S1 are established based on the structural characteristics of the mechanical equipment being diagnosed. Running signal set U={u1,u2,…,u m }, where u1, u2, ..., u m This represents different operating signals of the device, where m is the number of operating signals; establish the operating state set V={v1,v2,…,v…} n },v1,v2,…,v n This represents the running state, where n is the number of running states. The fault diagnosis model in S6 includes: a 2D-CNN channel, a 1D-CNN channel, and a fusion channel; The 2D-CNN channel includes an input layer (input2d), a two-dimensional convolutional layer (conv2d), a two-dimensional max pooling layer (max_pooling2d), a two-dimensional convolutional layer (conv2d_1), a two-dimensional max pooling layer (max_pooling2d_1), an attention mechanism module layer (se_attention_layer), a two-dimensional convolutional layer (conv2d_2), a two-dimensional max pooling layer (max_pooling2d_2), and a flatten layer. 1D-CNN includes an input layer (input1d), a one-dimensional convolutional layer (conv1d), a one-dimensional max pooling layer (max_pooling1d), a one-dimensional convolutional layer (conv1d_1), a one-dimensional max pooling layer (max_pooling1d_1), an attention layer (attention_layer), a one-dimensional convolutional layer (conv1d_2), a one-dimensional max pooling layer (max_pooling1d_2), and a flatten layer (flatten_1). The fusion channel includes a concatenate layer, a dropout layer, a feature layer, and an output layer. The time-sensitive running signal is processed to obtain a two-dimensional time-frequency graph, which is then stitched together and used as the input to the 2D-CNN channel input layer input2d; the running signal is stitched together and used as the input to the 1D-CNN channel input layer input1d; the 2D-CNN channel flatten layer flatten and the 1D-CNN channel flatten_1 layer flatten_1 are used as the input to the concatenate layer.

2. The mechanical equipment fault diagnosis method based on multi-signal fusion according to claim 1, characterized in that, The noise reduction process in S3 specifically includes: S3-1: The acquired running signal is decomposed using fully adaptive noise complete set empirical mode decomposition to obtain n intrinsic mode functions and one residual signal. The n intrinsic mode functions are represented as IMF1, IMF2, ..., IMF1. n The residual signal is represented as r; S3-2: The correlation coefficients between the n intrinsic mode functions and the original signal are calculated using the Pearson correlation coefficient; the calculation method of the Pearson correlation coefficient is shown in formula (1). （1） In the formula: x represents the original signal, IMF i Let i represent the IMF component of level i; according to formula (1), when the original signal and the IMF component are negatively correlated, the correlation coefficient is negative, and the higher the linear correlation, the larger the correlation coefficient is. S3-3: Signal reconstruction to obtain the denoised signal; When the correlation coefficient between the intrinsic mode function and the original signal When the value is less than a set threshold, the intrinsic mode function is identified as noise and removed. When the correlation coefficient between the intrinsic mode function and the original signal When the intrinsic mode function is greater than or equal to a set threshold, it is identified as the signal generated and is retained. The retained intrinsic mode function is then reconstructed with the residual signal to form a denoised signal.

3. The mechanical equipment fault diagnosis method based on multi-signal fusion according to claim 1, characterized in that, The specific construction method of the sample set in S5 is as follows: S5-1: Determine the sample data based on whether the operating signal is sensitive to time changes; S5-2: Process the time-sensitive operating signal samples to obtain a two-dimensional time-frequency diagram. Use CEEMDAN to decompose the time-sensitive operating signal, and then use Hilbert transform (HT) to perform Hilbert transform on each IMF component after CEEMDAN decomposition to generate a two-dimensional time-frequency diagram.

4. The mechanical equipment fault diagnosis method based on multi-signal fusion according to any one of claims 1-3, characterized in that, The diagnostic model for S7 is established by the following steps: S7-1: Divide the sample set into a training set and a test set according to a certain ratio, with the ratio of the number of samples in the training set to the number of samples in the test set being 7:3; S7-2: Set training parameters. The Adam optimizer was used to update the network parameters. The Epoch, batch size and learning rate of the Adam algorithm during model training were 60, 64 and 0.001, respectively. S7-3: Model training. The fault diagnosis model parameters are initialized in a random manner. The training set samples are input into the fault diagnosis model in batches for model training. The test samples are input into the trained fault diagnosis model for testing. The fault diagnosis accuracy is output to determine the feasibility of the proposed model. After training, the trained fault diagnosis model is saved.

5. The mechanical equipment fault diagnosis method based on multi-signal fusion according to claim 1, characterized in that, In S8, the fault diagnosis specifically involves inputting the set of operating signals U to be diagnosed into the trained fault diagnosis model, performing the diagnosis, and outputting the diagnosis results.

Images