Deep neural network electric shovel mechanical fault identification method based on multivariate feature extraction

Abstract

The invention discloses a deep neural network electric shovel mechanical fault identification method based on multivariate feature extraction. The method comprises the steps of data acquisition and preprocessing; constructing a deep neural network for fault detection; realizing a life prediction technology; and constructing an intelligent operation and maintenance decision of the mechanical electric excavator. According to the method, the fault detection process of the electric shovel of the mechanical electric excavator is decomposed into a plurality of subtasks, and training and optimization of a deep learning model are carried out for time domain, frequency domain and time-frequency domain features respectively. According to the method, the model can more accurately learn and identify fault modes in different feature spaces, so that the accuracy of fault detection is improved.

Description

Technical Field

[0001] The present invention relates to the field of electric shovel fault identification in deep neural networks, and specifically to a method for identifying mechanical faults of electric shovels based on deep neural networks with multi-feature extraction. Background Art

[0002] The fault detection of electric shovels is a key link in the maintenance of mechanical electric excavators, which directly affects the operation efficiency and safety of the equipment. In the traditional fault detection process, technicians usually rely on experience and regular inspections to identify potential faults. This process is not only time-consuming but also highly subjective, making it difficult to achieve accurate and real-time fault warnings. To improve the accuracy and efficiency of fault detection, researchers have begun to explore automated and intelligent fault detection technologies, among which the convolutional neural network (CNN) model based on deep learning is particularly prominent.

[0003] Early electric shovel fault detection mainly relied on manual feature extraction, such as time-domain features, frequency-domain features, and time-frequency domain features. These features need to be extracted by experts based on experience and complex signal processing techniques. This method is limited by the knowledge and experience of experts on the one hand, and is inefficient in processing large-scale data and prone to missing key information on the other hand. With the development of deep learning technology, researchers have begun to use convolutional neural networks to automatically learn features directly from raw data. This method not only improves the efficiency of feature extraction but also can capture complex and subtle patterns in the data, thus enhancing the performance of fault detection.

[0004] In recent years, the application of deep learning technology in the field of fault detection has made remarkable progress. Researchers have constructed fault detection models based on CNN, such as adaptive CNN models, BiLSTM models, CNN-LSTM models, and Transformer models, etc., to directly learn fault patterns from time-domain, frequency-domain, and time-frequency domain features. These models can automatically adjust parameters and optimize the network structure to adapt to different data features and fault types. However, although these methods have achieved certain results in global prediction accuracy, there are still challenges in the accuracy and error in local regions (such as the detection of specific fault patterns).

[0005] To further improve the accuracy and robustness of electric shovel fault detection, researchers have begun to explore more complex network structures and optimization strategies, such as adaptively adjusting the CNN model structure, hybrid models combining long short-term memory networks (LSTM) and convolutional neural networks (CNN), and using the Transformer architecture for feature extraction and fault prediction. These advanced technologies not only improve the accuracy of fault detection but also reduce the false alarm rate, providing strong technical support for the intelligent operation and maintenance of mechanical electric excavators. Summary of the Invention

[0006] In view of the technical challenges in the field of fault detection for mechanical and electric excavators, the present invention proposes a method for identifying mechanical faults of electric shovels based on a convolutional neural network with multi - feature extraction. This method aims to improve the accuracy of fault detection, reduce the false alarm rate, and provide a scientific basis for the intelligent operation and maintenance of equipment.

[0007] To achieve the above - mentioned invention purpose, the present invention adopts the following technical solutions:

[0008] A method for identifying mechanical faults of electric shovels based on a deep neural network with multi - feature extraction, including:

[0009] Construct an operation and maintenance decision database containing quadruples for mechanical and electric excavators; each piece of data in the operation and maintenance decision database covers fault features, fault components, fault causes, and maintenance measures; the fault features are the corresponding vibration feature data when the corresponding fault components fail due to the corresponding fault causes.

[0010] Obtain the vibration signal and rotation speed signal during the operation of the mechanical and electric excavator, and perform feature extraction on the collected vibration signal and rotation speed signal to obtain vibration feature data and rotation speed feature data.

[0011] Input the obtained vibration feature data into the trained electric shovel fault detection model to predict the fault state of the mechanical and electric excavator at the current moment.

[0012] When the prediction result output by the electric shovel fault detection model indicates that the mechanical and electric excavator has a fault at the current moment, query the operation and maintenance decision database to obtain the fault components, fault causes, and maintenance measures of the mechanical and electric excavator at the current moment.

[0013] When the prediction result output by the electric shovel fault detection model indicates that the mechanical and electric excavator has no fault at the current moment, input the obtained vibration feature data and rotation speed feature data into the remaining life prediction model to judge the time to the next fault, and then query the operation and maintenance decision database to obtain the fault components, fault causes, and maintenance measures corresponding to the next fault.

[0014] The electric shovel fault detection model is constructed based on a deep learning architecture and is trained using a fault sample data set.

[0015] The remaining life prediction model is constructed based on a self - attention mechanism and is trained using a remaining life sample data set.

[0016] Preferably, the sample features of the fault sample data set are the vibration feature data obtained after feature extraction of the vibration signals collected during the operation of the mechanical and electric excavator, and the sample labels are fault labels.

[0017] The sample features of the lifespan sample dataset are the vibration feature data and rotational speed feature data obtained after feature extraction from the vibration signals and rotational speed signals collected during the operation of a mechanical electric excavator, and the sample label is the possible fault occurrence time.

[0018] Preferably, for the feature extraction of the vibration signal and rotational speed signal, specifically: perform Fourier transform on the obtained vibration signal and rotational speed signal, and extract time-domain, frequency-domain, and time-frequency domain features, so as to obtain the corresponding feature data.

[0019] Preferably, the electric shovel fault detection model is an adaptive CNN model, BiLSTM model, CNN-LSTM model, or Transformer model.

[0020] Preferably, when the electric shovel fault detection model is an adaptive CNN model, it includes a fault detection input layer, convolution layer 1, batch normalization layer 1, pooling layer 1, convolution layer 2, batch normalization layer 2, pooling layer 2, fully connected layer 1, Dropout layer, and fully connected layer 2, where:

[0021] The input of the fault detection input layer is the vibration feature data;

[0022] Convolution layer 1 extracts features from the input vibration feature data through convolution operations;

[0023] Batch normalization layer 1 performs batch normalization on the output result of convolution layer 1;

[0024] Pooling layer 1 performs max pooling on the batch-normalized result;

[0025] Convolution layer 2 continues to perform convolution operations based on the pooling result of the previous layer;

[0026] Batch normalization layer 2 performs batch normalization on the output result of convolution layer 2;

[0027] Pooling layer 2 then performs max pooling on the result after batch normalization by batch normalization layer 2;

[0028] Fully connected layer 1 flattens the output of pooling layer 2 and then performs a fully connected operation;

[0029] Dropout layer performs a random inactivation operation on the output of fully connected layer 1;

[0030] Fully connected layer 2 performs another fully connected operation on the output of Dropout layer, and the output of fully connected layer 2 is the output of the electric shovel fault detection model.

[0031] Preferably, during the training process of the electric shovel fault detection model, a cross-entropy loss function and an Adam optimizer are used.

[0032] Preferably, the life prediction model is a Transformer life prediction model based on the self-attention mechanism, including a life prediction input layer, an embedding layer, a position encoding layer, a multi-head attention layer 1, a feed-forward neural network layer 1, a layer normalization layer 1, a multi-head attention layer 2, a feed-forward neural network layer 2, a layer normalization layer 2, and a fully connected layer, where:

[0033] The input of the life prediction input layer is vibration feature data and rotational speed feature data;

[0034] The embedding layer is used to perform an embedding operation on the input vibration feature data and rotational speed feature data, and map them to a specific vector space dimension;

[0035] The position encoding layer is used to perform a position encoding operation on the result embedded by the embedding layer;

[0036] The multi-head attention layer 1 is used to perform a multi-head attention operation on the result after adding the position encoding to integrate the feature information related to each position of the input data;

[0037] The feed-forward neural network layer 1 is used to perform feed-forward neural network processing on the output of the multi-head attention layer 1;

[0038] The layer normalization layer 1 is used to perform layer normalization operation on the output of the feed-forward neural network layer 1;

[0039] The multi-head attention layer 2 continues to perform a multi-head attention operation based on the result of the previous layer normalization;

[0040] The feed-forward neural network layer 2 is used to perform feed-forward neural network processing on the output of the multi-head attention layer 2;

[0041] The layer normalization layer 2 is used to perform layer normalization operation on the output of the feed-forward neural network layer 2;

[0042] The fully connected layer is used to perform a fully connected operation on the output of the layer normalization layer 2;

[0043] The output of the fully connected layer is the output of the life prediction model.

[0044] Another technical object of the present invention is to provide an electronic device, including a memory, a processor, and a computer program stored on the memory and running on the processor, and the computer program runs to execute the above-mentioned deep neural network-based electric shovel mechanical fault identification method based on multi-feature extraction.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0046] 1. Automatic Feature Learning: Through the application of deep learning technology, the present invention realizes automatic feature learning from raw data, reduces the dependence on expert experience and manual feature extraction, and improves the accuracy and efficiency of fault detection.

[0047] 2. Model Flexibility and Robustness: The model of the present invention can adapt to different data features and fault types. By adaptively adjusting parameters and network structures, the flexibility and robustness of fault detection are improved.

[0048] 3. Improving Prediction Performance: By introducing the Transformer architecture, the present invention can more effectively process non-stationary signals, capture temporal features, and improve the performance of fault detection.

[0049] 4. Intelligent Operation and Maintenance Support: Through the application of deep learning technology, the present invention realizes fast and accurate detection of electric shovel faults, provides strong technical support for the intelligent operation and maintenance of mechanical electric excavators, helps reduce maintenance costs, and improves the safety and reliability of equipment operation.

[0050] 5. Reducing Maintenance Costs: The method of the present invention can reduce maintenance costs caused by false alarms and missed alarms. Through accurate fault prediction, optimize the maintenance plan, reduce unexpected downtime, and improve production efficiency.

[0051] 6. Improving Equipment Safety: By real-time monitoring and accurately predicting potential faults, the present invention helps to take timely measures to avoid safety accidents caused by equipment failures and ensure the safety of operators and equipment. Brief Description of the Drawings

[0052] Figure 1 is the flowchart of the method of the present invention; Detailed Embodiment

[0053] In this embodiment, a method for identifying mechanical faults of an electric shovel based on a convolutional neural network with multi-feature extraction aims to solve the problems of insufficient accuracy and low efficiency in electric shovel fault detection in the prior art. This method combines time-domain, frequency-domain, and time-frequency-domain features, and uses a deep learning model to automatically extract features and perform fault prediction to improve the performance of fault detection.

[0054] Specifically, the method for identifying mechanical faults of an electric shovel based on a deep neural network with multi-feature extraction according to the present invention includes:

[0055] Construct an operation and maintenance decision database containing quadruples for a mechanical electric excavator electric shovel; each piece of data in the operation and maintenance decision database covers fault features, faulty components, fault causes, and maintenance measures; the fault features are the corresponding vibration feature data when the corresponding faulty components fail due to the corresponding fault causes.

[0056] Obtain the vibration signal and rotational speed signal of a mechanical electric excavator during operation, and perform feature extraction on the collected vibration signal and rotational speed signal to obtain vibration feature data and rotational speed feature data.

[0057] The feature extraction of the vibration signal and rotational speed signal is specifically as follows: perform Fourier transform on the obtained vibration signal and rotational speed signal, and extract time-domain, frequency-domain, and time-frequency domain features, so as to obtain the corresponding feature data.

[0058] Input the obtained vibration feature data into the trained electric shovel fault detection model to predict the fault state of the mechanical electric excavator at the current moment;

[0059] When the prediction result output by the electric shovel fault detection model indicates that the mechanical electric excavator has a fault at the current moment, query the operation and maintenance decision database to obtain the faulty components, fault causes, and maintenance measures of the mechanical electric excavator at the current moment;

[0060] When the prediction result output by the electric shovel fault detection model indicates that the mechanical electric excavator has no fault at the current moment, input the obtained vibration feature data and rotational speed feature data into the remaining life prediction model to judge the time to the next fault, and then query the operation and maintenance decision database to obtain the faulty components, fault causes, and maintenance measures corresponding to the next fault;

[0061] The electric shovel fault detection model is constructed based on a deep learning architecture and is trained using a fault sample data set;

[0062] The remaining life prediction model is constructed based on the self-attention mechanism and is trained using a remaining life sample data set.

[0063] The method for identifying mechanical faults of an electric shovel based on deep neural network with multi-feature extraction according to the present invention generally includes the following four parts:

[0064] The first part, data collection and preprocessing;

[0065] Step 1.1: Collect the vibration signal (one-dimensional data, with the horizontal axis representing time and the vertical axis representing acceleration) of the mechanical electric excavator during operation. Normalize these signal data to eliminate the influence of dimension;

[0066] Step 1.2: Perform Fourier transform on the collected signal data, and extract time-domain, frequency-domain, and time-frequency domain features. These features will be used as the input of the convolutional neural network model for training and predicting faults.

[0067] (1) Time-domain features

[0068] In the time-domain signal representation, f = (x1,…,x D) is a representation of discrete-time domain signals. Among them, f represents the entire time-domain signal, which is an ordered combination composed of D elements. x i (i = 1, 2, …, D) represents the value of the signal at the i-th sample point (or moment). The actual meaning of this value depends on the type of the signal. If it is an electric current signal, x i is the current intensity at this moment. D represents the length of the signal, that is, the number of sample points, which reflects the fineness of the description of the signal in the time dimension. The larger D is, the more sufficient the description of the signal's details over time. And although this representation does not directly give time information, the time corresponding to each sample point can be determined by assuming a fixed sampling interval, thereby completely depicting the signal in the time domain. Specifically, these time-domain features cover many important indicators such as mean, variance, root mean square value, peak value, peak-to-peak value, kurtosis, skewness, etc. The mean reflects the average level of the signal in the time domain; the variance reflects the degree of dispersion of the signal relative to the mean; the root mean square value comprehensively considers the energy size of the signal; the peak value and the peak-to-peak value respectively show the highest value that the signal can reach and the difference range between the highest value and the lowest value; kurtosis is sensitive to the impact components in the signal and can be used to detect abnormal fluctuations in the signal; skewness can characterize the asymmetry of the signal distribution. These different time-domain features depict the characteristics of the signal in the time domain from multiple dimensions, providing a very valuable data basis and strong support for subsequent in-depth signal analysis, fault diagnosis, and system performance evaluation. The specific mathematical statistical formulas are as follows:

[0069] Mean:

[0070] Root mean square value:

[0071] Variance:

[0072] Absolute mean value:

[0073] Root amplitude:

[0074] Skewness:

[0075] Kurtosis:

[0076] (2) Frequency-domain features

[0077] Mainly through the method of fast Fourier transform (FFT), various frequency-domain features are statistically analyzed as follows:

[0078] 1. Peak frequency

[0079] Let the discrete frequency sequence be \(f = \{f_1, f_2, \ldots, f\) n \}\), and the corresponding amplitude sequence be \(A = \{A_1, A_2, \ldots, A\) n \}\). To obtain the peak frequency, first, we need to find the index \(k\) corresponding to the maximum value in the amplitude sequence \(A\), that is:

[0080]

[0081] Here, \(\text{argmax}\) represents the index value of the independent variable that makes the function value reach the maximum. Then, the peak frequency \(f\) peak (can be expressed as:

[0082] \(f\) peak \(= f\) k

[0083] Combined, the calculation formula for the peak frequency can be written as:

[0084]

[0085] 2. Peak - amplitude

[0086] For the peak amplitude, we can directly find the maximum value in the amplitude sequence \(A\). Let the peak amplitude be \(A\) peak , and its mathematical formula is expressed as:

[0087] \(A\) peak \(= \max(A_1, A_2, \ldots, A\) n )

[0088] 3. Total - energy

[0089] According to the logic of calculating the total energy, first, square each element in the amplitude sequence \(A\), and then sum them to get the total energy. Let the total energy be \(E\) total , and its calculation formula is as follows:

[0090]

[0091] \(A\) peak \(= \max(A_1, A_2, \ldots, A\) n ) and the formula for calculating the total energy of

[0092] (2) Time - frequency domain features

[0093] Mainly through the method of short - time Fourier transform (STFT), various time - frequency domain features are statistically analyzed, specifically as follows:

[0094] Dominant frequency: freqs \(i\) i \(= f\)max_amplitude_idx' ;

[0095] wherein wherein is an element in the time-frequency matrix obtained after STFT. Here, i represents the frequency index, j represents the time segment index, and Zxx_mag ij represents the amplitude-related value at the j-th time segment and the i-th frequency component. The maximum amplitude of each time segment: amx_amplitude e k = amx i∈{1,2,…,m} (Zxx_m ing ); The total energy of each time segment:

[0096] Step 1.3: Mark the fault data and assign fault labels to each sample. These labels can be the true fault types based on historical maintenance records or the potential fault types preset through an expert system.

[0097] Part Two: Construct a convolutional neural network for fault detection:

[0098] Step 2.1: Select a suitable deep learning architecture, such as a CNN, BiLSTM, CNN-LSTM, or Transformer model, to construct an electric shovel fault detection model. Taking the CNN model as an example, construct a network structure including multiple convolutional layers, pooling layers, and fully connected layers.

[0099] Fault detection input layer: Let the input feature data be x0, and the dimension and other relevant information correspond to the requirements of the original input. The data form can be regarded as a vector form (in the case of one-dimensional data, the dimension can be extended and understood according to the specific application scenario).

[0100] Convolutional layer 1 (conv1): Extract features from the input x0 through a convolution operation, which can be expressed as:

[0101] x1 = f conv1 (x0)

[0102] where f conv1 represents a convolution operation with a kernel size of (3,), a stride of (1,), an input channel number of 1, an output channel number of 16, and a suitable padding (padding = (1,)). The dimension and other features of the output result x1 are correspondingly changed through this operation.

[0103] Batch normalization layer 1 (bn1): Perform batch normalization on the result x1 output by convolutional layer 1, and the mathematical expression is:

[0104] x2 = f bn1 (x1)

[0105] Here, f m1 represents the operation of normalizing the data according to the given parameters (eps = 1e - 05, momentum = 0.1, etc.). The purpose is to accelerate training and improve model stability, etc. After processing, x2 maintains appropriate dimensionality and data distribution characteristics.

[0106] Pooling layer 1 (pool1): Perform max - pooling operation on the result x2 after batch normalization:

[0107] x3 = f pool1 1 (x2)

[0108] f pooll is the max - pooling operation with a kernel size of 2, a stride of 2, no additional padding, etc. After pooling, the data dimensionality changes again to obtain x3

[0109] Convolutional layer 2 (conv2): Continue with the convolution operation based on the previous pooling result x3:

[0110] x4 = f convi (x3)

[0111] f conv22 is a convolution operation with a kernel size of (3,), a stride of (1,), 16 input channels, 32 output channels, and corresponding padding. The features of the output x4 change accordingly.

[0112] Batch normalization layer 2 (bn2): Perform batch normalization on the output x4 of convolutional layer 2:

[0113]

[0114] f bm2 Normalize the data according to the set parameters (eps = 1e - 05, momentum = 0.1, etc.) to obtain x5. Pooling layer 2 (pool2): Then perform max - pooling on the batch - normalized x5:

[0115] x6 = f pool2 (x5)

[0116] f pool2 is the max - pooling operation (with kernel size of 2, stride of 2, etc., related parameter settings). After processing, x6 is obtained, and the data dimensionality changes further.

[0117] Fully - connected layer 1 (fc1): Flatten the output x6 of pooling layer 2 and then perform a fully - connected operation:

[0118] x7 = f fc1 (x6)

[0119] Here, f fc1is a linear transformation that transforms the input feature dimension from the dimension after being processed by the previous layers to an output of 64 dimensions, which includes operations such as a suitable weight matrix and bias (bias = True), and the output is x7.

[0120] Dropout layer (dropout): Randomly deactivates the output x7 of the fully connected layer 1 to obtain x8, in order to prevent overfitting:

[0121] x8 = f dropout (x7)

[0122] Fully connected layer 2 (fc2): Finally, perform another fully connected operation:

[0123] x9 = f fc2 (x8)

[0124] f fc2 is a linear transformation that converts the input 64 - dimensional features into a 5 - dimensional output (corresponding to the number of fault classifications, etc., and bias = True means there is a bias participating in the operation). The final output x9 is the fault classification result predicted by the model (with a dimension of 5, and the value of each dimension corresponds to the probability of a different fault category).

[0125] Step 2.2: Input the pre - processed feature data into the model and obtain a fault detection model through training. During the training process, use the cross - entropy loss function and the Adam optimizer to minimize the difference between the model prediction and the actual fault labels.

[0126] The loss function is selected for the case of binary classification or multi - classification (when converting a multi - classification problem into multiple binary classification problems). It combines the Sigmoid function and the binary cross - entropy loss, which is convenient for directly outputting the unactivated original prediction values for training optimization. The optimizer uses optim.Adam, and the learning rate (lr) is set to 0.001. The Adam optimizer combines advantages such as adaptive learning rate adjustment, and can dynamically adjust the learning rate according to the gradient conditions of different parameters to accelerate the convergence of the model.

[0127] Step 2.3: Validate and test the model to evaluate the accuracy and robustness of the model. Use independent validation sets and test sets to evaluate the model to ensure that the model has good generalization ability on unseen data.

[0128] Part Three: Build a life prediction model to achieve life prediction.

[0129] The life prediction of equipment in large electric shovel excavators essentially belongs to a supervised regression learning task. Its sample features are derived from the vibration signals and rotational speed signals provided by each sensor, while the sample labels are the moments when faults may occur in the equipment where each sensor is located in the future.

[0130] The present invention adopts a Transformer lifespan prediction model based on the self-attention mechanism, specifically:

[0131] Lifespan prediction input layer: Let the input feature data be x0, and the dimension and other relevant information correspond to the requirements of the original input. The data form can be regarded as a vector form.

[0132] Embedding Layer: Perform an embedding operation on the input feature data x0, and map it to a specific vector space dimension. The mathematical representation is:

[0133] x1 = f_mbedding(x0)

[0134] Among them, f_embedding is an embedding function. According to the pre-set embedding dimension (for example, set to 256 dimensions), the input data is dimensionally transformed using the learned embedding matrix, so that x1 has the embedding dimension characteristics suitable for subsequent processing and can better capture the hidden information in the data.

[0135] Position Encoding Layer: Since the Transformer architecture itself is not sensitive to sequence order information, position encoding needs to be added. Perform a position encoding operation on the embedded result x1:

[0136] x2 = f_position_encoding(x1)

[0137] The f_position_encoding function adds position information to the vector of each position according to specific rules (generating different position encoding values using sine and cosine functions). The processed x2 contains both data feature representations and incorporated position information, ensuring that the model can distinguish the order of elements at different positions in the input data, which is particularly important for processing time series-related features involved in lifespan prediction.

[0138] Multi-HeadAttention Layer 1: Perform a multi-head attention operation on x2 after adding position encoding:

[0139] x3 = f_multihead_attention1(x2)

[0140] Here, f_multihead_attention1 represents the multi-head attention mechanism, which consists of multiple parallel self-attention heads (the number of heads is set to 6). Each self-attention head calculates the degree of association between different positions of the input data, and then combines and linearly transforms the results of these heads. The output x3 integrates the feature information of the correlations between different positions of the input data, and its dimensions and other features change accordingly, enabling the potential relationships in the data to be explored from a more comprehensive perspective and providing a richer feature basis for life prediction.

[0141] Feed Forward Neural Network Layer 1: Next, perform feed-forward neural network processing on the output x3 of the multi-head attention layer 1:

[0142] x4 = f_feed_forward1(x3)

[0143] f_feed_forwar1 is a feed-forward neural network containing two fully connected layers (the first linear transformation expands the dimension, and the second restores the dimension). An appropriate activation function (ReLU) can be used in the middle to introduce non-linearity. After processing by this layer, the features of the output x4 of the feed-forward neural network layer 1 are further transformed, and the dimension may also change according to the network structure, thereby extracting more advanced and abstract features, which helps in the recognition of patterns related to the equipment life.

[0144] Layer Normalization Layer 1: Perform layer normalization operation on the output x4 of the feed-forward neural network layer 1:

[0145] x5 = f_layer_norm1(x4)

[0146] f_layer_norm1(x4) normalizes the data according to the given parameters (mean, variance, etc. related settings). The purpose is to stabilize the gradient propagation during network training, keep the input data distribution of each layer relatively stable, and the processed output x5 maintains appropriate dimensions and data distribution characteristics, ensuring the efficiency and stability of model training.

[0147] Multi-Head Attention Layer 2: Continue with the multi-head attention operation, based on the result x5 normalized by the previous layer:

[0148] x6 = f_multihead_attention2(x5)

[0149] f_multihead_attention2 is also a multi-head attention mechanism, whose function is similar to that of the multi-head attention layer 1, but the parameters and inputs are different. It captures the correlations between different positions of the input data again. The output x6 integrates new correlation feature information, and its dimensions and so on change accordingly, further strengthening the mining and integration of data features in order to more accurately predict the equipment life.

[0150] Feed Forward Neural Network Layer 2: Perform a feed forward neural network process on the output x6 of the multi-head attention layer 2:

[0151] x7 = f_feed_forward2(x6)

[0152] f_feed_forward2 is also a feed forward neural network with a specific structure, similar to the function of the feed forward neural network layer 1, used to further transform the data features. After passing through this layer, the dimensions and features of the output x7 change again, extracting feature information that can better reflect the changing trend of the equipment life.

[0153] Layer Normalization Layer 2: Perform a layer normalization operation on the output x7 of the feed forward neural network layer 2:

[0154] x8 = f_layer_norm2(x7)

[0155] f_layer_norm2 normalizes the data according to the corresponding parameters, making the data distribution more conducive to subsequent processing and model training, and the output is denoted as x8.

[0156] Fully Connected Layer: Perform a fully connected operation on the output x8 of the layer normalization layer 2:

[0157] x 9 = f_fully_connected(x 8 )

[0158] Here, f_fully_connected is a linear transformation that transforms the input feature dimension from the dimension after being processed by the previous layers to an output of 1 dimension (representing the predicted life value), including appropriate weight matrix and bias (bias = True) operations, etc. The final output x9 is the result of the model's prediction of the equipment life.

[0159] Part Four: Construct the Intelligent Operation and Maintenance Decision-making of the Mechanical Electric Excavator

[0160] The essence of the operation and maintenance decision-making advice based on fault diagnosis and life prediction is actually a database construction task, and the specific process is as follows:

[0161] First, sort out the equipment, components and fault types that the sensors can detect; second, for each type of fault of each type of component, determine the corresponding maintenance measures through research; third, match the corresponding fault characteristics (sensor data) to each fault with the sensor data with fault labels (true or predicted); fourth, construct an operation and maintenance decision-making database containing quadruples, and each piece of data covers fault characteristics, fault components, fault causes and maintenance measures; fifth, in actual use, input the sensor data. If the fault diagnosis model determines that there is a fault, obtain the fault component, fault cause and maintenance measure by querying the database. If it is determined that there is no fault, use the life prediction model to judge the time to the next fault, and then query the database to obtain the fault component, fault cause and maintenance measure corresponding to the next fault.

[0162] Therefore, the present invention has the following characteristics:

[0163] 1. Deep learning model construction: The present invention proposes a series of fault detection models based on deep learning, including an adaptive CNN model, a BiLSTM model, a CNN-LSTM model and a Transformer model. These models can automatically learn complex features directly from the original data without relying on traditional manual feature extraction methods, thereby improving the efficiency and accuracy of feature extraction.

[0164] 2. Adaptive model adjustment: The models of the present invention can adaptively adjust parameters and network structures according to different data characteristics and fault types. For example, the adaptive CNN model can automatically adjust the convolution kernel size, stride and padding according to different time-domain features, frequency-domain features and time-frequency domain features to achieve the optimal fault detection effect.

[0165] 3. Hybrid model application: The present invention also proposes a hybrid model combining a long short-term memory network (LSTM) and a convolutional neural network (CNN). This hybrid model can simultaneously capture the long-term dependence relationship of time series data and the local features of spatial data, improving the accuracy of fault detection.

[0166] 4. Transformer architecture: The present invention introduces the Transformer architecture, which can effectively process sequence data and capture time series features, providing a new technical means for the fault detection of electric shovels. The Transformer model can handle long-distance dependence problems through the self-attention mechanism and is suitable for the fault detection of non-stationary signals.

Claims

1. A method for identifying mechanical faults of an electric shovel based on multi - feature extraction by a deep neural network, characterized in that Including: Construct an operation and maintenance decision database containing quadruples for mechanical electric excavators; Each piece of data in the operation and maintenance decision database covers fault characteristics, faulty components, fault causes, and maintenance measures; The fault characteristics are the corresponding vibration characteristic data when the corresponding faulty component fails due to the corresponding fault cause; Obtain the vibration signal and rotational speed signal of the mechanical electric excavator during operation, and perform feature extraction on the collected vibration signal and rotational speed signal to obtain vibration characteristic data and rotational speed characteristic data; Input the obtained vibration characteristic data into the trained electric shovel fault detection model to predict the fault state of the mechanical electric excavator at the current moment; When the prediction result output by the electric shovel fault detection model indicates that the mechanical electric excavator has a fault at the current moment, query the operation and maintenance decision database to obtain the faulty component, fault cause, and maintenance measures of the mechanical electric excavator at the current moment; when the prediction result output by the electric shovel fault detection model indicates that the mechanical electric excavator has no fault at the current moment, input the obtained vibration characteristic data and rotational speed characteristic data into the remaining life prediction model to judge the time to the next fault, and then query the operation and maintenance decision database to obtain the faulty component, fault cause, and maintenance measures corresponding to the next fault; The electric shovel fault detection model is constructed based on a deep learning architecture and is trained using a fault sample data set; The remaining life prediction model is constructed based on the self-attention mechanism and is trained using a remaining life sample data set.

2. The method for identifying mechanical faults of an electric shovel based on multi-feature extraction of a deep neural network according to claim 1, characterized in that The sample characteristics of the fault sample data set are the vibration characteristic data obtained after feature extraction of the vibration signal collected during the operation of the mechanical electric excavator, and the sample label is the fault label; The sample characteristics of the remaining life sample data set are the vibration characteristic data and rotational speed characteristic data obtained after feature extraction of the vibration signal and rotational speed signal collected during the operation of the mechanical electric excavator, and the sample label is the possible fault occurrence time.

3. The method for identifying the mechanical faults of an electric shovel based on a deep neural network with multi-feature extraction according to claim 2, wherein The feature extraction of the vibration signal and rotational speed signal is specifically: perform Fourier transform on the obtained vibration signal and rotational speed signal, and extract time domain, frequency domain, and time-frequency domain characteristics to obtain the corresponding feature data.

4. The method for identifying mechanical faults of an electric shovel based on deep neural network with multi - feature extraction according to claim 3, wherein, The electric shovel fault detection model is an adaptive CNN model, BiLSTM model, CNN-LSTM model, or Transformer model.

5. The method for identifying mechanical faults of an electric shovel based on deep neural network with multi-feature extraction according to claim 4, wherein When the electric shovel fault detection model is an adaptive CNN model, it includes a fault detection input layer, convolutional layer 1, batch normalization layer 1, pooling layer 1, convolutional layer 2, batch normalization layer 2, pooling layer 2, fully connected layer 1, Dropout layer, and fully connected layer 2, where: The input of the fault detection input layer is the vibration characteristic data; Convolutional layer 1 extracts features from the input vibration characteristic data through convolution operations; Batch normalization layer 1 performs batch normalization on the result output by convolutional layer 1; Pooling layer 1 performs max pooling on the batch-normalized result; Convolutional layer 2 continues to perform convolution operations based on the previous pooling result; Batch normalization layer 2 performs batch normalization on the result output by convolutional layer 2; Pooling layer 2 then performs max pooling on the result after batch normalization by batch normalization layer 2; The fully connected layer 1 flattens the output of the pooling layer 2 and then performs a fully connected operation; The Dropout layer performs a random inactivation operation on the output of the fully connected layer 1; The fully connected layer 2 performs another fully connected operation on the output of the Dropout layer, and the output of the fully connected layer 2 is the output of the electric shovel fault detection model.

6. The method for identifying mechanical faults of an electric shovel based on a deep neural network with multi-feature extraction according to claim 4, characterized in that During the training process of the electric shovel fault detection model, the cross-entropy loss function and the Adam optimizer are used.

7. The method for identifying the mechanical faults of an electric shovel based on a deep neural network with multi-feature extraction according to claim 5, characterized in that The life prediction model is a Transformer life prediction model based on the self-attention mechanism, including a life prediction input layer, an embedding layer, a position encoding layer, a multi-head attention layer 1, a feed-forward neural network layer 1, a layer normalization layer 1, a multi-head attention layer 2, a feed-forward neural network layer 2, a layer normalization layer 2, and a fully connected layer, where: The input of the life prediction input layer is vibration feature data and rotational speed feature data; The embedding layer is used to perform an embedding operation on the input vibration feature data and rotational speed feature data, and map them to a specific vector space dimension; The position encoding layer is used to perform a position encoding operation on the result embedded by the embedding layer; The multi-head attention layer 1 is used to perform a multi-head attention operation on the result after adding the position encoding to integrate the feature information of the input data among different positions; The feed-forward neural network layer 1 is used to perform feed-forward neural network processing on the output of the multi-head attention layer 1; The layer normalization layer 1 is used to perform layer normalization on the output of the feed-forward neural network layer 1; The multi-head attention layer 2 continues to perform a multi-head attention operation based on the result of the previous layer normalization; The feed-forward neural network layer 2 is used to perform feed-forward neural network processing on the output of the multi-head attention layer 2; The layer normalization layer 2 is used to perform layer normalization on the output of the feed-forward neural network layer 2; The fully connected layer is used to perform a fully connected operation on the output of the layer normalization layer 2; The output of the fully connected layer is the output of the life prediction model.

8. An electronic device, including a memory, a processor, and a computer program stored on the memory and running on the processor, where the computer program runs to execute the method for identifying mechanical faults of an electric shovel based on a deep neural network with multi-feature extraction according to any one of claims 1 to 7.

Images