Asynchronous motor fault diagnosis method and system based on CNN-LSTM hybrid deep learning model

Abstract

The invention relates to the technical field of asynchronous motor diagnosis, in particular to an asynchronous motor fault diagnosis method based on a CNN-LSTM mixed deep learning model, and the method comprises the steps: carrying out the multi-source sensor signal collection, data preprocessing and entropy weight method fusion, and parallel CNN feature extraction and LSTM time sequence modeling. According to the method, the end-to-end intelligent fault diagnosis is realized, the dependence on expert experience is remarkably reduced, the diagnosis precision and robustness are improved, a feasible path is provided for industrial predictive maintenance, and the maintenance cost and personal errors are reduced. The modular design of the system supports online deployment, outputs fault types and probabilities in real time, and assists engineers in making decisions quickly. The method is not only suitable for asynchronous motors, but also can be expanded to fault diagnosis of other rotating machines, and has wide application prospects and technical popularization values.

Description

Technical Field

[0001] This invention relates to the field of asynchronous motor diagnostic technology, and more specifically, to an asynchronous motor fault diagnosis method based on a CNN-LSTM hybrid deep learning model. Background Technology

[0002] Traditional asynchronous motor fault diagnosis methods heavily rely on human experience for feature extraction, such as requiring experts to manually analyze vibration spectra or current signal characteristics. This is not only inefficient but also susceptible to subjective factors, leading to poor diagnostic consistency. In complex industrial environments, motor operating states exhibit strong nonlinearity and multi-physics coupling characteristics, making it difficult for traditional methods to capture deep fault modes, often resulting in insufficient diagnostic accuracy and a high false alarm rate. Furthermore, existing technologies are mostly based on single sensor data (such as using only vibration signals), failing to fully utilize the complementarity of multi-source heterogeneous information (such as current, voltage, and temperature), thus limiting the comprehensiveness and reliability of fault detection.

[0003] With the development of sensor technology, multi-source data acquisition has become possible, but how to effectively fuse these data and achieve automated analysis remains a bottleneck. Existing fusion methods, such as simple weighted averaging or principal component analysis, often ignore the spatiotemporal correlation between data and are sensitive to noise, resulting in poor robustness in real-time applications. While deep learning models such as CNN or LSTM have made progress in some areas, a single model struggles to simultaneously handle spatial feature extraction and time series modeling, limiting its performance in motor fault diagnosis, especially its ability to identify early, weak, or complex faults.

[0004] Therefore, the industry urgently needs an intelligent method capable of automatically processing multi-source sensor data, fusing spatial and temporal features, and achieving high-precision end-to-end diagnostics. This invention addresses these issues by proposing a solution based on multi-sensor information fusion and a CNN-LSTM hybrid model, aiming to overcome the limitations of traditional technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, such as reliance on manual feature extraction, insufficient diagnostic accuracy in nonlinear, strongly coupled systems, and difficulty in effectively utilizing multi-source heterogeneous sensor data, this invention proposes a fault diagnosis method and system for asynchronous motors using a CNN-LSTM hybrid deep learning model. This method achieves end-to-end intelligent fault diagnosis by combining multi-source sensor signal acquisition, data preprocessing and entropy weighting fusion, and parallel CNN feature extraction with LSTM temporal modeling. This significantly reduces reliance on expert experience and improves diagnostic accuracy and robustness.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for diagnosing faults in an asynchronous motor includes the following steps: The multi-source sensor signal acquisition step is used to acquire various types of sensor signals from the asynchronous motor, including current, voltage, temperature, and vibration signals. The data preprocessing and multi-source information fusion steps are used to denoise, normalize, and slice the sensor signals, and generate a fused signal using an information fusion algorithm. The fault diagnosis step of the CNN-LSTM hybrid model is used to automatically extract spatial and temporal features from the fused signal and output fault classification results; The model training step is used to train the CNN-LSTM hybrid model through an optimization algorithm to minimize the loss function; The fault diagnosis and assessment steps are used to perform online or offline diagnosis based on the trained model and to calculate performance metrics.

[0007] Preferably, in the multi-source sensor signal acquisition step, the sensors include a current sensor, a voltage sensor, a temperature sensor, and a vibration sensor; and the signal acquisition is performed under various speed and fault conditions, including normal, stator fault, rotor fault, bearing inner ring fault, bearing outer ring fault, rolling element fault, and mixed fault.

[0008] Preferably, the normalization process in the data preprocessing and multi-source information fusion steps adopts the minimum-maximum normalization algorithm, and its calculation formula is as follows:

[0009] in, It is the raw data. These are the minimum and maximum values ​​of the data in that channel, respectively. It is the normalized data, ranging from [0, 1].

[0010] Preferably, the multi-source information fusion step employs the entropy weighting method for data layer fusion, and includes the following sub-steps:

[0011] in, It is the first The first sensor The proportion of each data value in the sensor data sequence after normalization.

[0012] Then, the weight of each sensor is determined using the entropy weight method. The smaller the information entropy, the greater the variability of the sensor's data, the more information it provides, and therefore the greater its weight should be. Weight of each sensor The calculation is as follows:

[0013] Finally, the weighted average yields the fused signal. :

[0014] in, For the number of sensors, Let be the signal value of the j-th sensor.

[0015] Preferably, the CNN-LSTM hybrid model includes multiple convolutional neural network (CNN) sub-modules configured in parallel and a long short-term memory network (LSTM) module; the CNN sub-modules process different fused signals respectively, and each sub-module contains four sets of stacked convolutional layers and pooling layers for extracting spatial features; the LSTM module receives the feature sequence output by the CNN sub-modules for learning temporal dependencies.

[0016] Preferably, the convolutional layer operation formula in the CNN submodule is as follows:

[0017] Where X is the input sequence, K is the convolution kernel, F is the kernel size, and Y is the output feature map; the ReLU activation function is used after the convolutional layer, defined as... The pooling layer uses average pooling, and its operation formula is as follows:

[0018] Where W is the width of the pooling window. For input values.

[0019] Preferably, the LSTM module includes a forget gate, an input gate, and an output gate, and its calculation formula is as follows: Forgotten Gate:

[0020] Input gate

[0021]

[0022] Output gate:

[0023]

[0024] in It is the Sigmoid activation function. Here is the weight matrix for the forget gate. This is the bias value of the input gate.

[0025] Preferably, the model training step uses the Adam optimization algorithm, with a batch size of 64 and 800 iterations; the loss function is the classification cross-entropy, and its formula is:

[0026] M is the sample size in a batch, and K is the number of fault categories. It is an indicator function. Calculated using the Softmax function: .

[0027] Preferably, the fault diagnosis and evaluation step includes calculating accuracy, recall, and precision metrics, as shown in the following formulas: Accuracy:

[0028] Recall rate:

[0029] Accuracy:

[0030] in This represents the number of correctly classified samples A; Indicates the remaining number of misclassifications; This represents the number of misclassified samples A. This represents the number of remaining samples that were correctly classified.

[0031] The present invention also provides an asynchronous motor fault diagnosis system for implementing the method described in any of the above embodiments, comprising: A multi-sensor data acquisition module is used to collect current, voltage, temperature, and vibration signals of the asynchronous motor in real time. The data preprocessing and fusion module is used to denoise, normalize, slice, and fuse the acquired signals using the entropy weight method. The CNN-LSTM fault diagnosis model module is used for feature extraction and fault classification using a hybrid CNN-LSTM model. The diagnostic results output and display module is used to visualize the fault type, probability, and confidence level.

[0032] The present invention has the following beneficial effects: This invention achieves high-precision, automated fault diagnosis of asynchronous motors by integrating multi-sensor information with a CNN-LSTM hybrid deep learning model, significantly improving detection efficiency and reliability. This is attributed to the CNN module's ability to automatically extract local spatial features of signals and the LSTM module's long-term modeling of time dependencies, thus comprehensively capturing the dynamic features of the fault. Experiments show that this method maintains stable performance under various speeds and fault conditions, demonstrating its strong robustness.

[0033] Another key benefit of this invention is the reduction of reliance on manual feature engineering, enabling end-to-end intelligent diagnosis. By using entropy weighting for multi-source data fusion and automatically weighting information from different sensors, the representativeness of features is enhanced. Simultaneously, the model training employs the Adam optimizer and early stopping method to avoid overfitting and improve generalization ability. Compared to CNN-LSTM models using only a single vibration signal, this invention significantly improves accuracy, highlighting the core advantage of multi-source fusion in providing complementary physical information, making it suitable for complex industrial environments.

[0034] Furthermore, this method provides a feasible path for predictive maintenance in industry, reducing maintenance costs and human error. The modular system design supports online deployment, providing real-time output of fault types and probabilities to assist engineers in rapid decision-making. This invention is not only applicable to asynchronous motors but can also be extended to fault diagnosis of other rotating machinery, possessing broad application prospects and technological promotion value. Attached Figure Description

[0035] Figure 1 A schematic diagram of the overall architecture of an asynchronous motor fault diagnosis method based on a CNN-LSTM hybrid deep learning model provided in an embodiment of the present invention; Figure 2 This is a block diagram of the overall structure of an asynchronous motor fault diagnosis system based on a CNN-LSTM hybrid deep learning model in an embodiment of the present invention. Figure 3 This is a schematic diagram of a multi-sensor fusion CNN-LSTM neural network model in an embodiment of the present invention; Figure 4 This is a schematic diagram of the LSTM unit structure in an embodiment of the present invention; Figure 5 Here are the accuracy (a) and loss value (b) of the model training set and validation set in this embodiment of the invention; Figure 6 This is a schematic diagram of the confusion matrix in an embodiment of the present invention; Figure 7 This is an example image of a two-dimensional clustered scatter plot in an embodiment of the present invention. Detailed Implementation

[0036] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0037] The present invention will now be further described with reference to the accompanying drawings and specific embodiments. The embodiments of the present invention provide a method and system for fault diagnosis of asynchronous motors based on multi-sensor information fusion and a hybrid deep learning model of convolutional neural network-long short-term memory network (CNN-LSTM). This method aims to solve the problems of traditional fault diagnosis techniques, such as strong dependence on manual feature extraction, insufficient diagnostic accuracy in nonlinear strongly coupled systems, and difficulty in effectively utilizing multi-source heterogeneous sensor data.

[0038] like Figure 1 As shown, an asynchronous motor fault diagnosis method based on multi-sensor information fusion and a hybrid deep learning model of convolutional neural network-long short-term memory network (CNN-LSTM) mainly includes the following steps: S100: Multi-source sensor signal acquisition.

[0042] Specifically, various types of sensors are deployed at key locations of the asynchronous motor to comprehensively monitor its operating status. In a preferred embodiment, the deployed sensors include: Current sensor: Used to collect stator current signals (unit: A) of phase A, phase B, and phase C.

[0043] Voltage sensor: Used to collect stator voltage signals (unit: V) of phase A, phase B, and phase C.

[0044] Temperature sensor: Used to collect temperature signals (unit: °C) of phase A, phase B, and phase C windings, as well as temperature signals of the stator core.

[0045] Vibration sensor: Used to collect vibration signals (unit: m / s) in the horizontal and vertical directions of the front bearing. 2 ).

[0046] Specifically, the data acquisition parameters can be set as follows: sampling frequency of 500Hz and continuous sampling time of 40 seconds. Accordingly, each sensor channel can acquire 500 Hz * 40 s = 20,000 data points. To simulate different motor operating conditions, the experiment can be conducted at various motor speeds (e.g., 1510 r / min, 2250 r / min, 2840 r / min) and various fault states (e.g., normal operation, stator fault, rotor fault, bearing inner ring fault, bearing outer ring fault, rolling element fault, and mixed fault).

[0047] S200: Data preprocessing and multi-source information fusion.

[0048] This step is crucial for ensuring data quality and improving model performance, and it is further subdivided into the following sub-steps: S210: Data preprocessing.

[0049] Specifically, the preprocessing of the raw signals acquired in step S100 mainly includes: Denoising: Digital filters (such as low-pass filters and wavelet thresholding) are used to remove high-frequency noise and power frequency interference from the signal while retaining the effective fault characteristic frequency band.

[0050] Normalization: To eliminate the influence of different physical dimensions and numerical ranges and accelerate model convergence, the denoised data is normalized. A commonly used method is min-max normalization, and its calculation formula is as follows:

[0051] in, It is the raw data. These are the minimum and maximum values ​​of the data in that channel, respectively. It is the normalized data, ranging from [0, 1].

[0052] Data slicing: To construct samples suitable for training deep learning models, long-term time-series data is cut into fixed-length sample segments. For example, 600 consecutive data points are taken as one sample. For 20,000 points collected over 40 seconds, approximately 33 samples can be obtained (20000 / 600≈33.33, discarding the remainder). Data is collected at different speeds for each fault state, and a sufficient number of samples can be obtained by summarizing. For example, a dataset containing 7000 samples can be constructed (7 fault states * 1000 samples / state).

[0053] S220: Fusion of similar information from multiple sources.

[0054] Specifically, multiple signals from similar sensors are fused into a more representative and robust composite signal. This aims to reduce data dimensionality and enhance the representational power of features. This invention employs the entropy weighting method for data layer fusion, and the specific process is as follows: Suppose there are m sensors of the same type (such as three-phase current sensors, m=3), and each sensor collects data at n time points.

[0055] First, calculate the information entropy of each sensor signal. For the signal from the j-th sensor, its information entropy... The calculation formula is:

[0056] in, It is the first The first sensor The proportion of each data value in the sensor data sequence after normalization.

[0057] Then, the weight of each sensor is determined using the entropy weight method. The smaller the information entropy, the greater the variability of the sensor's data, the more information it provides, and therefore the greater its weight should be. Weight of each sensor The calculation is as follows:

[0058] Finally, the weighted average yields the fused signal. :

[0059] This method fuses the three-phase stator current into a "fused stator current," the three-phase stator voltage into a "fused stator voltage," the three-phase winding temperature into a "fused winding temperature," and the bearing vibrations in both directions into a "pre-fusion bearing vibration." Finally, these, along with the original "fused stator core temperature" signal, constitute six sets of multi-source heterogeneous information, which serve as input to the subsequent CNN-LSTM model.

[0060] S300: Construct a CNN-LSTM fault diagnosis model based on multi-channel input.

[0061] This step is the core of the method of this invention, and the model structure is as follows: Figure 3 As shown, this model employs a parallel-serial hybrid architecture, fully leveraging the advantages of CNNs in spatial feature extraction and LSTMs in time series modeling.

[0062] S310: Convolutional Neural Network (CNN) module construction.

[0063] The CNN module consists of 6 sub-models with identical structures and independent parameters. Each sub-model processes a set of multi-source heterogeneous information generated in step S220.

[0064] S311: Input layer.

[0065] The input layer of each CNN sub-model receives a preprocessed and fused one-dimensional temporal signal. The shape of the input data is [batch size, number of channels, sequence length]. Since it is a univariate temporal signal, the number of channels is usually 1. For example, in a batch processing, the input shape can be represented as [64, 1, 600], indicating a batch size of 64 and a sample length of 600.

[0066] S312: Convolutional Layer.

[0067] Convolutional layers are the core of CNNs, used to automatically extract local features from the input signal. Their operation is achieved by performing convolution operations between multiple trainable convolutional kernels (filters) and the input data.

[0068] Convolution operation: For one-dimensional convolution, given an input sequence X (of length W) and a convolution kernel K (of size F), the convolution operation outputs the feature map Y with the th... The formula for calculating each element is:

[0069] In practical applications, padding (P) and stride (S) also need to be considered. The length N of the output feature map of the convolutional layer is determined by the following formula:

[0070] in, This indicates rounding down. In one embodiment of the invention, the parameters of the first convolutional layer can be set as follows: kernel size F=8, padding P=0, stride S=2. Then the output length N=(600-8+2*0) / 2+1=297.

[0071] Weight sharing: The same convolutional kernel slides across the entire input sequence and shares parameters, which greatly reduces the number of model parameters.

[0072] Activation Function: After convolution, a non-linear activation function, such as ReLU (Rectified Linear Unit), is typically applied to enhance the model's non-linear expressive power. The ReLU function is defined as follows:

[0073] Using ReLU can effectively alleviate the vanishing gradient problem and accelerate the training process.

[0074] S313: Pooling Layer.

[0075] Pooling layers typically follow convolutional layers to downsample feature maps, reducing data dimensionality and computational cost while enhancing feature invariance. This invention primarily employs average pooling.

[0076] Average pooling operation: Within a local region (such as a window of size w), the average of all values ​​is taken as the output. Its calculation formula can be expressed as:

[0077] in, Indicates the first Layer The first feature map The output of each neuron Indicates the l-th layer The first feature map The input of each neuron, This is the width of the pooled window. For example, setting the pooled window size. If the input length is 5 and the step size is S=2, then the output length of the feature map with an input length of 297 after pooling is approximately (297-5) / 2+1=147.

[0078] 314: Multi-layer convolution and pooling stacking.

[0079] To extract deeper and more abstract features, each CNN sub-model in this embodiment of the invention contains four stacked sets of convolutional layers and pooling layers. Specific parameters are shown in the table below:

[0080] By stacking multiple layers, the model can progressively extract rich spatial features from the original signal, ranging from low to high levels.

[0081] S320: Construction of Long Short-Term Memory (LSTM) Network Module.

[0082] The LSTM module is used to learn the long-term dependencies of multi-source heterogeneous information features over time. Its core unit structure is as follows: Figure 4 As shown.

[0083] S321: LSTM cell principle and calculation.

[0084] LSTM introduces the concept of a memory cell, using a state This allows LSTM to transmit and store information, making it more flexible in forgetting and retaining information. The formula for calculating the memory unit is:

[0085] in Represents the current cellular memory state. The activation value for the forget gate. Indicates the state at the previous moment; The output of the input gate; This represents the state variables of the candidate elements. The input signal for the current cell; This is the output signal of the previous cell.

[0086] (2) Gate of Oblivion The forgetting gate, as the main control gate for forgotten information, combines the input signal X with the output H of the previous neuron. After nonlinear processing by the sigmoid activation function, the output f of the forgetting gate is obtained. The calculation formula for the forgetting gate is:

[0087] in It is the Sigmoid activation function. Here is the weight matrix for the forget gate. This is the bias value of the input gate.

[0088] (3) Input gate The input gate, acting as the gateway to updating information in the memory unit, is calculated using the following formula:

[0089]

[0090] in This represents the weight parameter matrix of the input gate. This is the forget gate bias value. The weight matrix for the subsequent selected vectors. Let represent the bias vector, and Tanh be the hyperbolic tangent activation function.

[0091] (4) Output gate The output gate, acting as the output gateway of the memory unit, is calculated using the following formula:

[0092]

[0093] in This indicates the result of the output gate. This represents the weight matrix of the subsequent vectors. This represents the bias vector. This indicates that the output value is hidden.

[0094] S322: LSTM layer configuration.

[0095] Specifically, the long feature sequence concatenated in step S315 (e.g., a sequence of length 1088 * 6 / after some recombination method) is input into the LSTM layer. The configuration parameters of the LSTM layer can include: input dimension (matching the feature vector dimension), hidden layer dimension (e.g., 110), and number of stacked layers (e.g., 1 layer). The hidden state of the LSTM layer at the last time step... Alternatively, the sum of all hidden states (such as the average) can be passed to subsequent classification layers as a contextual representation of the entire input sequence.

[0096] S330: Construction of the classification output layer.

[0097] Specifically, the high-level feature representation output by the LSTM module is fed into a fully connected layer, which maps the learned distributed features to the label space of the samples. Finally, the Softmax activation function is used to output the probability distribution belonging to each fault category. Assuming there are K fault categories (e.g., K=7), for the input sample... It belongs to the first class probability Given by the Softmax function:

[0098] in, This is the weight vector corresponding to the j-th class. The training objective of the model is to maximize the log probability of the correct class, i.e., minimize the categorical cross-entropy loss function:

[0099] Here, M is the sample size in a batch. It is an indicator function, when the sample Authentic Labels equal The value is 1 if it is active, and 0 otherwise.

[0100] S400: Model Training and Optimization.

[0101] This step aims to learn the optimal parameters of the model in a data-driven manner.

[0102] S410: Dataset partitioning.

[0103] The fully labeled dataset (e.g., 7000 samples) constructed in step S200 is randomly divided into training, validation, and test sets. A common partitioning ratio is 80% for training, 10% for validation, and 10% for testing. For example, out of 1000 samples for each fault class, 800 are used for training, 100 for validation, and 100 for testing.

[0104] S420: Training parameter settings.

[0105] Specifically, the following strategies are used to train the CNN-LSTM model: Optimizer: The Adam (Adaptive Moment Estimation) optimization algorithm is used, which combines the advantages of momentum method and RMSProp. It can adaptively adjust the learning rate and usually converges faster and performs well.

[0106] Learning rate: Set an initial learning rate (e.g., 0.001) and use a learning rate decay strategy.

[0107] Batch Size: Set to 64, meaning that 64 samples are used in each iteration to calculate the loss and update the gradient.

[0108] Number of iterations (Epochs): Set to 800, meaning the model will traverse the training set 800 times.

[0109] Regularization: Use Dropout in fully connected layers to randomly drop a portion of neurons (e.g., dropout rate set to 0.5) to prevent model overfitting.

[0110] Early stopping: Monitor the loss or accuracy on the validation set, and terminate training early when the validation set performance no longer improves over several consecutive epochs to avoid overfitting.

[0111] S430: Training process monitoring.

[0112] During training, the loss and accuracy curves on the training and validation sets are recorded and plotted in real time, such as... Figure 5 As shown in the figure. These curves provide a visual indication of whether the model has converged and whether there is overfitting or underfitting. In this embodiment of the invention, the model tends to stabilize after approximately 200 iterations and reaches its optimal performance after 800 iterations.

[0113] S500: Fault diagnosis and performance evaluation.

[0114] The trained CNN-LSTM model was used for fault diagnosis of actual asynchronous motors.

[0115] S510: Online / offline diagnostics.

[0116] Specifically, for the motor to be diagnosed, its multi-sensor data (or historical data) is collected in real time. After preprocessing and fusion in step S200, the data is input into the deployed CNN-LSTM model in step S300. After forward propagation, the model outputs a 7-dimensional probability vector, and the category with the highest probability is the fault type predicted by the model.

[0117] S520: Model evaluation index.

[0118] accuracy Recall (R) and precision (P) are common evaluation metrics for fault diagnosis based on deep learning. Precision is the ratio of the number of correctly classified samples to the total number of samples; recall is the ratio of correctly classified samples of a particular class to the total number of samples of that class, representing the ability to identify positive samples; precision is the ratio of the number of correctly classified samples of a particular class to the number of predicted samples, representing the ability to identify negative samples.

[0119]

[0120]

[0121]

[0122] in This represents the number of correctly classified samples A; Indicates the remaining number of misclassifications; This represents the number of misclassified samples A. This represents the number of remaining samples that were correctly classified.

[0123] S530: Comparative Experiment and Result Analysis.

[0124] To verify the superiority of this invention, extensive comparative experiments were conducted. The multi-sensor fusion CNN-LSTM model proposed in this invention was compared with the following benchmark models: 1D-CNN model: Only uses CNN modules (excluding LSTM), and the input is multi-sensor fusion data.

[0125] LSTM model: Only LSTM module is used (without CNN), and the input is multi-sensor fusion data.

[0126] CNN-LSTM model (single vibration signal): This method uses the hybrid model proposed in this invention, but only a single vibration signal is input instead of multi-sensor fusion data.

[0127] Table 1 shows the average results of each model after 5 runs on the independent test set.

[0128] Analysis conclusion: The multi-sensor fusion CNN-LSTM model of this invention achieved the highest diagnostic accuracy (99.53%) and the lowest loss value (0.032), significantly outperforming other comparative models.

[0129] Compared to 1D-CNN, the accuracy is improved by approximately 6.41%, demonstrating the effectiveness of introducing LSTM to capture temporal dependencies.

[0130] Compared to LSTM, the accuracy is improved by approximately 9.11%, demonstrating the advantage of CNN in automatically extracting spatial features.

[0131] Compared to a CNN-LSTM model using only a single vibration signal, the accuracy was significantly improved by approximately 28.39%, strongly demonstrating the crucial role of multi-source information fusion in comprehensive and accurate motor fault diagnosis. Multi-sensor data provides complementary information, enabling the model to capture fault characteristics from different physical dimensions, thereby greatly enhancing the model's discriminative ability.

[0132] also, Figure 6 and Figure 7 The model's classification performance was visualized, showing that various fault characteristics were clearly distinguished, further validating the model's excellent clustering and diagnostic performance.

[0133] The asynchronous motor fault diagnosis method based on multi-sensor fusion CNN-LSTM provided in the above embodiments of the present invention achieves high-precision and robust diagnosis of motor faults in complex industrial environments through innovative multi-source information fusion strategies and deep learning model architecture. This method reduces reliance on human expert experience, realizes end-to-end intelligent diagnosis, and provides reliable technical support for predictive maintenance of industrial equipment.

[0134] It should be noted that the specific parameters mentioned in the above embodiments (such as kernel size, number of layers, learning rate, batch size, etc.), number of sensor types, and fusion algorithms (entropy weight method is only one example) can be adjusted and optimized according to the actual application scenario and data. These adjustments and optimizations should all be included within the protection scope of this invention.

[0135] like Figure 2 As shown, an asynchronous motor fault diagnosis system based on a CNN-LSTM hybrid deep learning model mainly includes: The multi-sensor data acquisition module 101 is responsible for acquiring signals such as vibration, current, voltage, and temperature from the asynchronous motor in real time. Data preprocessing and fusion module 102: Cleans, normalizes, slices, and fuses multi-source information on the collected raw data; The CNN-LSTM fault diagnosis model 103 is responsible for automatically extracting spatial and temporal features from the fused data and classifying faults. And the diagnostic result output and display module 104, which visualizes the diagnostic results (such as fault type, probability, and confidence level) to the user.

[0136] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for diagnosing faults in an asynchronous motor, characterized in that, Includes the following steps: The multi-source sensor signal acquisition step is used to acquire various types of sensor signals from the asynchronous motor, including current, voltage, temperature, and vibration signals. The data preprocessing and multi-source information fusion steps are used to denoise, normalize, and slice the sensor signals, and generate a fused signal using an information fusion algorithm. The fault diagnosis step of the CNN-LSTM hybrid model is used to automatically extract spatial and temporal features from the fused signal and output fault classification results; The model training step is used to train the CNN-LSTM hybrid model through an optimization algorithm to minimize the loss function; The fault diagnosis and assessment steps are used to perform online or offline diagnosis based on the trained model and to calculate performance metrics.

2. The method according to claim 1, characterized in that, In the multi-source sensor signal acquisition step, the sensors include current sensors, voltage sensors, temperature sensors, and vibration sensors; and the signal acquisition is carried out under various speed and fault conditions, including normal, stator fault, rotor fault, bearing inner ring fault, bearing outer ring fault, rolling element fault, and mixed fault.

3. The method according to claim 1, characterized in that, The normalization process in the data preprocessing and multi-source information fusion steps adopts the minimum-maximum normalization algorithm, and its calculation formula is as follows: ； in, It is the raw data. These are the minimum and maximum values ​​of the data in that channel, respectively. It is the normalized data, ranging from [0, 1].

4. The method according to claim 1, characterized in that, The multi-source information fusion step employs the entropy weighting method for data layer fusion, and includes the following sub-steps: ； in, It is the first The first sensor The weight of each data value in the sensor data sequence is determined after normalization; then, the weight of each sensor is determined using the entropy weight method. The smaller the information entropy, the greater the variability of the sensor data, the more information it provides, and the greater its weight should be. Weight of each sensor The calculation is as follows: ； Finally, the weighted average yields the fused signal. : ； in, For the number of sensors, Let be the signal value of the j-th sensor.

5. The method according to claim 1, characterized in that, The CNN-LSTM hybrid model includes multiple convolutional neural network (CNN) sub-modules set up in parallel and a long short-term memory network (LSTM) module; the CNN sub-modules process different fused signals respectively, and each sub-module contains four sets of stacked convolutional layers and pooling layers for extracting spatial features; The LSTM module receives the feature sequence output by the CNN submodule and uses it to learn time dependencies.

6. The method according to claim 5, characterized in that, The convolutional layer operation formula in the CNN submodule is as follows: ； Where X is the input sequence, K is the convolution kernel, F is the kernel size, and Y is the output feature map; the ReLU activation function is used after the convolutional layer, defined as... The pooling layer uses average pooling, and its operation formula is as follows: ； Where W is the width of the pooling window. For input values.

7. The method according to claim 5, characterized in that, The LSTM module includes a forget gate, an input gate, and an output gate, and its calculation formula is as follows: Forgotten Gate: ； Input gate ； ； Output gate: ； ； in It is the Sigmoid activation function. Here is the weight matrix for the forget gate. This is the bias value of the input gate.

8. The method according to claim 1, characterized in that, The model training steps use the Adam optimization algorithm, with a batch size of 64 and 800 iterations; the loss function is the classification cross-entropy, and its formula is: ； M is the sample size in a batch, and K is the number of fault categories. It is an indicator function. Calculated using the Softmax function: 。 9. The method according to claim 1, characterized in that, The fault diagnosis and evaluation steps include calculating accuracy, recall, and precision metrics, as shown in the following formulas: Accuracy: ； Recall rate: ； Accuracy: ； in This represents the number of correctly classified samples A; Indicates the remaining number of misclassifications; This represents the number of misclassified samples A. This represents the number of remaining samples that were correctly classified.

10. An asynchronous motor fault diagnosis system, used to implement the method according to any one of claims 1-9, characterized in that, include: A multi-sensor data acquisition module is used to collect current, voltage, temperature, and vibration signals of the asynchronous motor in real time. The data preprocessing and fusion module is used to denoise, normalize, slice, and fuse the acquired signals using the entropy weight method. The CNN-LSTM fault diagnosis model module is used for feature extraction and fault classification using a hybrid CNN-LSTM model. The diagnostic results output and display module is used to visualize the fault type, probability, and confidence level.

Images