Intelligent fault diagnosis method based on multi-input neural network

Abstract

The invention discloses an intelligent fault diagnosis method based on a multi-input neural network. The method comprises the steps of data acquisition, preprocessing, feature extraction, model training and fault judgment. A variety of sensor signals are collected from the device, de-noised and normalized to generate high quality features. Time domain and frequency domain features are extracted in combination with methods such as wavelet transform, deep learning is performed on feature data by using a neural network, and a potential fault mode is identified. Through the trained model, the system can detect and classify faults in real time, output fault levels and diagnosis reports, and provide support for equipment maintenance. The method is suitable for a large-scale complex environment, has high diagnosis precision and quick response capability, effectively improves the management efficiency, and reduces the non-planned downtime.

Description

Technical Field

[0001] The present invention belongs to the field of industrial automation and equipment fault diagnosis, and in particular relates to an intelligent fault diagnosis method based on a multi-input neural network. Background Art

[0002] In industrial production, equipment fault diagnosis and health monitoring are crucial for ensuring production efficiency and safety. Traditional equipment fault diagnosis methods rely heavily on manual experience, rule libraries, or shallow models based on single features. These methods face numerous limitations in today's complex and volatile industrial environment, particularly in terms of real-time performance, accuracy, and reliability.

[0003] With the advancement of industrial automation and intelligence, equipment fault diagnosis technology has become an important means of ensuring stable production line operation and improving production efficiency. For critical equipment in production lines, such as fans, pumps, reducers, and rolling mills, timely fault diagnosis and preventive maintenance can significantly reduce equipment downtime and repair costs. However, traditional fault diagnosis methods still face significant challenges in terms of real-time performance, accuracy, and multi-source data fusion, making them difficult to meet the monitoring needs of modern industrial equipment.

[0004] Currently, traditional equipment fault diagnosis technologies mostly rely on rule-based algorithms or shallow machine learning methods, which often have the following limitations:

[0005] (1) Lack of real-time performance

[0006] In industrial environments, equipment failures often require a swift response to prevent further damage. However, existing fault diagnosis methods often rely on offline analysis, making real-time monitoring and feedback difficult. Traditional methods typically store collected fault data before processing, resulting in data processing lags and difficulty responding to rapidly changing equipment conditions. When equipment is operating under high load, the inability to monitor and provide timely feedback on fault information can lead to uncontrolled downtime, severely impacting production efficiency.

[0007] (2) Limitations of accuracy and false alarm rate

[0008] Traditional fault diagnosis methods often analyze single features, such as the time or frequency domain characteristics of vibration signals, without integrating multiple features. This results in low accuracy in fault mode identification. In noisy environments, single features are easily affected by interference, leading to high false positive rates and an inability to accurately distinguish normal signal fluctuations from actual fault signals. Furthermore, due to the complexity of equipment failure modes, existing methods exhibit low accuracy when identifying multi-category or compound faults, increasing diagnostic uncertainty and risk.

[0009] (3) Difficulty in fusing multi-source data

[0010] The operating status of modern industrial equipment can be monitored using a variety of sensors. These sensors can provide multidimensional information about equipment operation, such as vibration, temperature, and current, forming multidimensional features in both the time and frequency domains. Traditional methods typically rely on single-type data processing, making it difficult to effectively integrate and leverage the correlations between multi-source data. This inability to fully reflect the equipment's operating status leads to incomplete diagnostic results. Due to the lack of data fusion methods, existing methods struggle to obtain reliable fault identification results when faced with complex equipment failure modes and uncertain environments.

[0011] (4) Lack of intelligent diagnostic methods

[0012] With the rapid advancement of intelligent manufacturing, modern equipment fault diagnosis systems must not only possess high accuracy and low false alarm rates, but also possess adaptive and intelligent diagnostic capabilities. However, traditional rule-based fault diagnosis methods cannot automatically adapt to changing operating conditions and environments, lacking the ability to adaptively identify new fault modes. In practical applications, equipment operating environments are complex and ever-changing. Traditional methods require constant parameter adjustments, and diagnostic models lack self-learning capabilities, making them difficult to meet real-time and intelligent requirements. Furthermore, traditional methods cannot update diagnostic models in real time, preventing them from continuously optimizing diagnostic accuracy and reliability based on new data.

[0013] (5) Typical methods in the prior art

[0014] Existing equipment fault diagnosis methods can be mainly divided into the following categories:

[0015] Rule-based fault diagnosis: This approach relies on expert experience and a pre-set rule base to perform pattern matching and analysis on device status data. This approach is simple and easy to implement, but the rule base is expensive to maintain and prone to failure as the device operating environment changes.

[0016] Statistical model-based fault diagnosis: This method analyzes signals and detects faults by calculating statistical characteristics such as the signal's mean, standard deviation, and variance. This method can identify abnormal signals to a certain extent, but it exhibits low accuracy when faced with complex and diverse fault modes.

[0017] Fault diagnosis based on shallow machine learning: This includes methods such as support vector machines (SVMs) and random forests, which train classification models based on simple features. While these methods can effectively classify faults in specific scenarios, they lack the ability to integrate multidimensional features and are less adaptable to changes in the equipment's operating environment.

[0018] In summary, traditional equipment fault diagnosis methods are unable to meet the real-time, efficient, and intelligent diagnosis needs of modern industrial equipment. Therefore, developing an intelligent fault diagnosis method that combines multi-source data fusion and deep learning to cope with the complex and changing operating environment of industrial equipment has become a technical challenge in the current field. Summary of the Invention

[0019] To solve the above problems, the present invention discloses an intelligent fault diagnosis method based on a multi-input neural network. This method realizes real-time monitoring of the equipment operating status and fault diagnosis by integrating multi-dimensional features and deep learning algorithms, thereby improving the management efficiency and stability of industrial equipment. It is suitable for health monitoring of complex industrial equipment such as high-speed rotating equipment, variable speed and load equipment (such as fans, water pumps, reducers and rolling mills).

[0020] To achieve the above object, the technical solution of the present invention is as follows:

[0021] The intelligent fault diagnosis method based on multi-input neural network includes the following steps:

[0022] S1: Data Collection

[0023] According to the monitoring requirements of the target equipment, multi-source data of the equipment's operating status is collected through sensors, including vibration, current, temperature and other signals.

[0024] S11: Use a MEMS accelerometer to collect vibration signals from the equipment, setting the sampling frequency to 10 kHz to capture high-frequency fault characteristic signals;

[0025] S12: Transmit sensor data to the central processing unit in real time through RS485 and CAN bus interfaces to ensure data timeliness and transmission stability;

[0026] S13: Simultaneously collect the current and temperature signals of the device to increase the diversity of the device status characteristics and provide a comprehensive data foundation for subsequent multi-dimensional feature fusion.

[0027] Through the data collection step, the present invention can obtain multi-dimensional data of equipment operation in real time, providing high-quality data support for data preprocessing and feature extraction.

[0028] S2: Data Preprocessing

[0029] The collected raw data is processed by denoising, standardization, outlier detection and time alignment to improve the quality and consistency of the data.

[0030] S21: Denoising. Use wavelet filtering to denoise the collected vibration signal. Select the 'db4' wavelet basis, set the decomposition level to 5, and use "soft thresholding" to remove high-frequency noise.

[0031] S22: Data standardization. Standardize the data from different sensors, adjust the mean to 0 and the standard deviation to 1 to ensure the uniformity of multi-source data.

[0032] S23: Outlier detection. Set an outlier detection threshold for the normalized data and remove outliers that exceed the threshold to improve data stability and consistency.

[0033] S24: Time alignment. Synchronize the time of different sensor data to ensure the temporal consistency of multi-source signals, facilitating subsequent feature fusion.

[0034] S25: Data formatting. Format the preprocessed data into a tensor format suitable for neural network model input, ensuring a unified and clear data structure.

[0035] The data preprocessing step improves the quality and timeliness of multi-source data, providing high-quality data support for subsequent feature extraction and model training.

[0036] S3: Feature Extraction

[0037] According to the diversity of fault types and the complexity of equipment operating status, time domain features, frequency domain features and time-frequency domain features are extracted to comprehensively characterize the operating status of the equipment.

[0038] S31: Time Domain Feature Extraction. Calculates 10 features of the vibration signal, including mean, standard deviation, peak value, RMS value, kurtosis, skewness, margin factor, and form factor, to describe the signal's fluctuations in the time domain.

[0039] S32: Time-frequency domain feature extraction. Continuous wavelet transform (CWT) is used to generate a 32x32 feature matrix to capture the time-frequency changes of the signal and improve the ability to identify sudden faults. The mathematical expression of CWT is as follows:

[0040]

[0041] in:

[0042] x(t): original signal;

[0043] ψ(t): mother wavelet function;

[0044] a: scale parameter, used to adjust frequency resolution;

[0045] b: Translation parameter, used to adjust the time resolution.

[0046] The Morlet wavelet is chosen as the mother wavelet because it has a good balance between time resolution and frequency resolution in frequency analysis and is suitable for capturing the instantaneous frequency variation characteristics of vibration signals. By setting the center frequency ω0 = 5, the efficiency and stability of the wavelet transform are effectively guaranteed. The Morlet wavelet is defined as follows:

[0047]

[0048] S33: Feature screening. Use correlation analysis and information gain algorithms to screen key features, eliminate redundant features, and improve the efficiency and accuracy of the diagnostic model.

[0049] Through the comprehensive extraction of multi-dimensional features, the present invention can accurately characterize the operating status of the equipment and improve the effectiveness and robustness of the diagnosis model.

[0050] S4: Neural Network Model Training

[0051] Build and train multi-input neural network models to achieve joint learning of different features and accurate classification of failure modes.

[0052] S41: Model Architecture Design. Build a multi-input neural network model, setting the time domain and time-frequency domain features as independent input channels; the time domain channel is processed using a fully connected layer, and the time-frequency domain channel is processed using a convolutional layer and a pooling layer, as follows:

[0053] (1) Time domain feature processing:

[0054] Input time domain features (such as mean, standard deviation, peak, etc.) to the fully connected layer. The first fully connected layer contains 64 neurons and the activation function is ReLU; the second layer contains 32 neurons.

[0055] Use Batch Normalization to normalize the output of each layer to prevent gradient disappearance or explosion.

[0056] (2) Time-frequency domain feature processing:

[0057] Input a two-dimensional feature matrix (size 32x32) to the convolution layer. The first convolution kernel size is 3x3, the stride is 1, and the number of output channels is 32; then the feature size is reduced through the maximum pooling layer (2x2).

[0058] The second convolution layer has a kernel size of 5x5 and 64 output channels, followed by a fully connected layer for feature dimensionality reduction.

[0059] S42: Feature fusion. Use the Concatenate layer to fuse the time domain and time-frequency domain features to form a unified feature vector. Use a three-layer fully connected network to further process the fused features.

[0060] S43: Deep Learning. This method uses multiple fully connected layers to perform deep learning on the fused features, ultimately outputting the probability distribution of various fault types through a softmax activation function, enhancing the ability to identify complex fault patterns.

[0061] S44: Model optimization. Using the Adam optimizer and the cross-entropy loss function, the model is optimized through early stopping and K-fold cross-validation to ensure its generalization and robustness. The optimization process of the model training process is as follows:

[0062] (1) Adaptive learning rate:

[0063] The Adam optimizer was used, and the initial learning rate was set to 0.001. When the validation set loss did not decrease for three consecutive times, the learning rate was halved.

[0064] (2) Early stopping strategy:

[0065] During model training, the loss value of the validation set is monitored. If the loss value does not improve for 10 consecutive iterations, the training is terminated to prevent overfitting.

[0066] (3) Regularization processing:

[0067] Dropout (the dropout rate is set to 0.5) is added after each fully connected layer to randomly ignore some neurons to enhance the generalization ability of the model.

[0068] The neural network model training step provides the present invention with accurate fault diagnosis capabilities and can effectively identify complex fault modes.

[0069] S5: Fault diagnosis and feedback

[0070] Based on the trained model, the equipment operating status is judged for faults and a real-time diagnostic report is generated.

[0071] S51: Fault identification. Real-time data is input into the neural network model, and the probability distribution of each fault type is output through the softmax activation function;

[0072] S52: Confidence Assessment. Calculate the confidence level of the fault type based on the probability distribution of the model output to help users assess the severity of the fault.

[0073] S53: Fault classification. Determine the fault category based on the maximum probability and record it in the system diagnosis report;

[0074] S54: Report generation. The system analyzes the real-time diagnostic results and automatically generates a report based on historical fault data. The report includes:

[0075] (1) Fault type: The specific fault category diagnosed by the model, such as bearing wear, misalignment, imbalance, etc.

[0076] (2) Occurrence time: The fault triggering time accurate to the second, used to analyze the changing trend of the equipment operating status.

[0077] (3) Confidence: Based on the classification probability output of the neural network, it indicates the confidence of the model in the diagnosis result.

[0078] (4) Risk level: Quantitative classification (low, medium, high) based on confidence and severity of the fault type.

[0079] (5) Recommended treatment measures: Combined with the expert knowledge base, provide targeted maintenance suggestions for specific faults.

[0080] S55: Visual display. The user interface displays diagnostic results in real-time in the form of charts and graphs, and supports users to query historical data and trend analysis. The system supports flexible query and filtering of equipment operation data, including:

[0081] (1) Time range selection: Users can customize the time range (e.g., the past hour, 24 hours, or one week) to filter fault data or historical operation records.

[0082] (2) Multi-dimensional query: Supports filtering by device name, fault type (such as bearing fault, gear wear) or operating parameters (such as vibration signal, temperature), facilitating targeted analysis.

[0083] (3) Trend analysis: Generate trend charts based on historical data, including fault frequency, equipment health change trends, etc.

[0084] (4) Alarm data screening: The system supports querying historical records by alarm level (low, medium, high), helping users quickly locate high-priority issues.

[0085] (5) Export function: Users can export query results in Excel or CSV format for easy archiving or integration with third-party management systems.

[0086] S56: Enhanced system functions. To facilitate users to understand the equipment operating status and fault information in real time, the system has the following extended functions:

[0087] (1) Real-time status monitoring: Displays the current equipment status in the form of a data dashboard, including dynamic changes in real-time parameters such as vibration, current, and temperature.

[0088] (2) Fault alarm: When the operating status of the equipment exceeds the preset threshold, the system will automatically trigger an alarm and mark the corresponding equipment for quick processing.

[0089] (3) Historical Query: Provides query function for historical operation data. Users can filter data according to time range or fault type, and conduct comprehensive analysis in combination with statistical charts.

[0090] (4) Permission management: Differentiated data access rights are provided based on user roles, and only administrators can modify alarm rules and thresholds.

[0091] (5) Data protection: Supports single sign-on (SSO) and multi-factor authentication (MFA) to ensure data security.

[0092] Through fault diagnosis and feedback steps, users can understand the operating status of the equipment in real time, providing a basis for maintenance decisions.

[0093] Compared with the prior art, the present invention has the following beneficial effects:

[0094] (1) High real-time performance: This invention achieves real-time monitoring and diagnostic response of equipment operating status through a neural network model and multi-input feature fusion, significantly improving the timeliness of fault diagnosis. Compared with traditional offline analysis methods, this invention can quickly capture changes in equipment status and adapt to the real-time processing requirements of high-frequency data.

[0095] (2) High Accuracy and Low False Alarm Rate: This invention utilizes multi-dimensional feature fusion and deep learning models to improve the accuracy of complex fault pattern recognition. Compared with traditional single-feature diagnostic methods, this invention significantly reduces the likelihood of false alarms and missed alarms, and exhibits higher robustness, especially in noisy industrial environments.

[0096] (3) Multi-source data fusion capability: By jointly learning time-domain and time-frequency domain features, the present invention can effectively fuse multi-source data from different sensors, improving the comprehensiveness of diagnosis. Compared with traditional methods that rely solely on a single feature, the multi-input neural network architecture of the present invention is more adaptable and can capture the operating status of complex equipment, improving the comprehensiveness and accuracy of fault identification.

[0097] (4) Intelligence and automation: Traditional methods often require frequent parameter adjustments or rely on manual experience and judgment. The deep learning model of the present invention has adaptive learning capabilities and can automatically update the model based on new equipment data to adapt to changes in equipment operating conditions. It has good intelligent characteristics and significantly reduces the need for manual intervention. BRIEF DESCRIPTION OF THE DRAWINGS

[0098] Figure 1 This is an application scenario and system architecture diagram of the intelligent fault diagnosis method based on multi-input neural network of the present invention;

[0099] Figure 2 This is a system function logic structure diagram of the intelligent fault diagnosis method based on multi-input neural network of the present invention;

[0100] Figure 3 This is an algorithm flow chart of the intelligent fault diagnosis method based on multi-input neural network of the present invention;

[0101] Figure 4 This is a structural diagram of the multi-input neural network model of the present invention. DETAILED DESCRIPTION

[0102] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0103] Example 1

[0104] This example uses a Nanjing steel plant as an example, focusing on key production equipment including rolling mills, speed reducers, fans, and other high-speed rotating equipment. These devices often experience state changes such as vibration, temperature rise, and abnormal current flow during operation. The system of the present invention enables real-time monitoring and intelligent diagnosis of these state changes.

[0105] like Figure 1 As shown, the equipment status monitoring and fault diagnosis system of the present invention includes multiple data acquisition units, a central data processing unit, a neural network model module, and a user interface module, forming a complete equipment status monitoring and fault diagnosis solution. The specific implementation steps are as follows:

[0106] Step S1: Data acquisition

[0107] According to production needs, sensors are placed on the main production equipment of the steel plant to realize real-time data collection of the equipment.

[0108] S11: Sensor Arrangement

[0109] MEMS accelerometers, temperature sensors, and current sensors are deployed in bearings, transmissions, motors, and other areas of key equipment such as rolling mills, fans, and reducers. Each piece of equipment is equipped with at least one vibration sensor and one current sensor, and multiple sensors can be installed as necessary to capture comprehensive equipment status information.

[0110] S12: Data Sampling

[0111] The sensor's sampling frequency is set at 10kHz to ensure comprehensive capture of the high-speed equipment's operating status. Data is transmitted to the central data processing unit via RS485 and CAN buses. To protect data transmission from the equipment's high electromagnetic interference, the sensor interface utilizes an anti-interference design and is isolated from the control cabinet.

[0112] S13: Data Synchronization and Storage

[0113] After receiving the sensor signal, the data processing unit synchronizes the data and stores it in the database to ensure that all sensor data are recorded on the same time axis for subsequent data processing and analysis.

[0114] Step S2: Data preprocessing

[0115] The data preprocessing module performs denoising, standardization and outlier detection on the collected original signals to ensure data clarity and stability, and is suitable for complex industrial environments.

[0116] S21: Denoising

[0117] The complex production environment of a steel plant is prone to high-frequency noise. To address this, we use wavelet filtering to remove high-frequency noise from the vibration signal, selecting the 'db4' wavelet basis and setting a 5-layer decomposition. The denoised signal is more suitable for fault feature extraction.

[0118] S22: Data Standardization

[0119] All signals are processed in a standardized manner with a mean of 0 and a standard deviation of 1 to ensure that the numerical range of data from different sensors is consistent, which is convenient for subsequent feature extraction module processing.

[0120] S23: Outlier Detection

[0121] Set anomaly detection thresholds for sudden data changes. For example, when a vibration or current signal exceeds a preset upper limit, the data point is automatically marked as an outlier and removed from the normal data set. After removing the outlier, the data is re-smoothed to ensure data continuity.

[0122] Outlier detection uses the Box Plot Method based on statistical analysis. The specific steps are as follows:

[0123] (1) Calculate the upper and lower quartiles Q1 and Q3 and the interquartile range IQR of the data using the following formula:

[0124] IQR=Q3–Q1

[0125] (2) Define the upper and lower limits of outliers:

[0126] Lower bound = Q1 - 1.5 × IQR

[0127] Upper bound = Q3 + 1.5 × IQR

[0128] (3) Eliminate data points that are not within this range to ensure the continuity and consistency of the data.

[0129] Step S3: Feature extraction

[0130] Based on the preprocessed data, time domain, frequency domain and time-frequency domain features are extracted to fully characterize the equipment operation status.

[0131] S31: Time Domain Feature Extraction

[0132] Extract time-domain features of device signals, including mean, standard deviation, peak value, kurtosis, skewness, and other indicators. These features can reflect the vibration and current fluctuations of the device under different operating conditions and help distinguish between normal and abnormal operating conditions.

[0133] S32: Time-frequency domain feature extraction

[0134] To address the complexity of vibration signals, we use the continuous wavelet transform (CWT) to convert time series signals into two-dimensional time-frequency plots and generate a 32x32 feature matrix to capture the signal's time-frequency variations. This time-frequency feature matrix is ​​more suitable for deep learning model training and helps identify sudden and complex faults.

[0135] Step S4: Neural network model training and deployment

[0136] The neural network model training module is responsible for learning and classifying multi-dimensional feature data. The model structure is as follows: Figure 3 shown.

[0137] S41: Model Architecture Design

[0138] We use Keras's Functional API to build a multi-input neural network model, which includes independent input channels for time-domain features and time-frequency domain features. Convolutional and pooling layers are used to extract time-frequency features, while the time-domain feature input channels are deep-learned through fully connected layers.

[0139] S42: Model Training

[0140] The model was trained using the Adam optimizer and the cross-entropy loss function, and K-fold cross-validation was employed to improve its generalization capabilities. The training data included historical and simulated data for multiple failure modes, ensuring that the model could identify common failure modes (bearing inner and outer race failures, rolling element failures, misalignment, imbalance, base looseness, cage failure, and impeller failure).

[0141] S43: Model Validation and Optimization

[0142] After model training, performance testing is performed using a validation set. Model performance is measured using metrics such as accuracy, recall, and F1 score. Model hyperparameters are adjusted based on the results to ensure robustness and high accuracy across various device failures.

[0143] S44: Model Deployment

[0144] The model is deployed to the central control system of the steel plant, which receives sensor data through a real-time interface and outputs fault diagnosis results, ensuring that the system can achieve real-time diagnosis in the production environment.

[0145] Step S5: Real-time fault diagnosis and feedback

[0146] In the actual production environment of a steel plant, the neural network model will receive sensor data in real time and output diagnostic results, and fault feedback will be displayed through the user interface module.

[0147] S51: Real-time data processing

[0148] Sensor data is received in real time and pre-processed before being fed into a neural network model. The model determines the device status based on the input data and provides a diagnosis of normal or faulty conditions.

[0149] S52: Fault classification and confidence output

[0150] The model outputs a probability distribution of fault types, using a softmax function to calculate fault confidence. The specific fault type is determined based on the maximum probability, and confidence information is output in the fault report to help on-site personnel assess the severity of the fault.

[0151] S53: Fault report generation and feedback

[0152] The generated fault report includes fault type, occurrence time, confidence level, recommended handling method, etc. The report content is displayed on the user interface, and serious faults can be promptly fed back to the operator through the alarm system.

[0153] S54: User Visualization Interface

[0154] The user interface provides real-time data visualization and fault history data query functions, helping steel plant maintenance personnel understand the health status of equipment and optimize maintenance management decisions.

[0155] Step S6: System adaptive optimization

[0156] In order to respond to changes in equipment operating conditions and new failure modes, the system adaptive optimization module can automatically update the model based on new data to ensure the efficiency and accuracy of long-term diagnosis.

[0157] S61: Model update mechanism

[0158] After detecting a new failure mode, the system adds the new data to the training set and improves its ability to recognize new patterns by regularly retraining the model.

[0159] S62: Performance Monitoring and Alarming

[0160] The system regularly checks the prediction accuracy of the model. When it finds that the fault recognition rate decreases or the false alarm rate increases, it will issue a warning signal to trigger the model optimization process.

[0161] It should be noted that the above content merely illustrates the technical idea of ​​the present invention and cannot be used to limit the scope of protection of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications all fall within the scope of protection of the claims of the present invention.

Claims

1. Based on the intelligent fault diagnosis method of multi-input neural network, an equipment status monitoring model algorithm system is constructed, which is characterized by: The steps include: S1: Collect equipment operation status data from equipment sensors, including vibration signals and current signals, and transmit the collected data to the central processing unit in real time; S2: Preprocess the collected data, including denoising, data alignment and outlier detection, and standardize the data format for subsequent analysis; S3: Extract the time domain features and time-frequency domain features of the device signal. The time domain feature extraction includes 10 features including mean, standard deviation, maximum value, minimum value, peak value, root mean square value, peak value, skewness, waveform factor and margin factor, and use wavelet transform to obtain time-frequency domain features; S4: Build a multi-input neural network model, perform feature fusion learning on time domain and time-frequency domain features through neural network, and train the model to identify multiple fault modes; S5: Predict the equipment operating status based on the trained neural network model, determine the fault type and generate a fault report.

2. The intelligent fault diagnosis method based on multi-input neural network according to claim 1 is characterized in that: In the step S1, the data is collected using a MEMS acceleration sensor, the device vibration is collected at a sampling frequency of 10 kHz, and the data is transmitted through an RS485 or CAN bus interface.

3. The intelligent fault diagnosis method based on multi-input neural network according to claim 1 is characterized in that: The step S2 comprises the following sub-steps: S21: Use wavelet filter to denoise the collected signal, select 'db4' wavelet basis, 5-layer decomposition, and use the "soft threshold" method as the threshold to eliminate high-frequency noise in the signal; S22: Data standardization: adjust the mean of the data to 0 and the standard deviation to 1 to ensure the consistency of different data dimensions; S23: Outlier detection, automatically detecting and removing abnormal data points by setting thresholds to ensure data continuity and consistency; S24: Time alignment processing to ensure synchronization of multi-source signals on the time axis for subsequent analysis; S25: Data formatting, converting the preprocessed data into a tensor format suitable for neural network input.

4. The intelligent fault diagnosis method based on multi-input neural network according to claim 1 is characterized in that: The step S3 specifically includes the extraction of time domain features and time-frequency domain features, wherein: a) Time domain feature extraction includes 10 features, including mean, standard deviation, maximum, minimum, peak, RMS value, peak, skewness, waveform factor and margin factor, which are used to describe the overall fluctuation characteristics of the signal; b) Time-frequency domain feature extraction uses continuous wavelet transform (CWT) to obtain a 32x32 feature matrix, which is used to analyze the changes in the frequency components of the signal over time.

5. The intelligent fault diagnosis method based on multi-input neural network according to claim 1 is characterized in that: The step S4 specifically includes the following sub-steps: S41: Construct a multi-input neural network model. The model structure includes a time domain feature input layer and a time-frequency domain feature input layer. Each input layer is connected to a convolution layer, a pooling layer, and a batch normalization layer. S42: Integrate the features of different input channels in the feature fusion layer to perform joint learning of multi-dimensional features; S43: Use the activation function ReLU to introduce nonlinear characteristics, and add a Dropout layer after the convolution layer to prevent overfitting; S44: Perform deep feature extraction and classification on features through the fully connected layer, and finally output the prediction results.

6. The intelligent fault diagnosis method based on multi-input neural network according to claim 5 is characterized in that: The neural network model adopts the Adam optimizer and the cross entropy loss function. The early stopping strategy and K-fold cross validation are used in model training to ensure the generalization ability and diagnostic accuracy of the model.

7. The intelligent fault diagnosis method based on multi-input neural network according to claim 1 is characterized in that: The step S5 specifically comprises the following steps: S51: The model forward propagates the input data and uses the softmax activation function to calculate the probability distribution of each fault type; S52: Determine the fault category according to the maximum probability of the fault type, and output the corresponding confidence level; S53: Fault severity assessment, determining the risk level of the fault based on historical fault data and confidence level; S54: The fault classification results are stored in a database and a fault report is generated for historical analysis and trend prediction.

8. The intelligent fault diagnosis method based on multi-input neural network according to claim 7 is characterized in that: The fault report generated in step S54 includes the fault type, occurrence time, confidence level, risk level and recommended treatment measures, and is displayed in the form of charts and reports through the user interface.

9. The intelligent fault diagnosis method based on multi-input neural network according to claim 8 is characterized in that: The fault report supports user-defined time range query, including real-time data, historical data, fault statistics and trend analysis to support equipment maintenance and management decisions.

10. The intelligent fault diagnosis method based on multi-input neural network according to claim 1 is characterized in that: The system provides a user interaction interface module, including real-time monitoring, fault alarm, history query and user authority management functions, and supports single sign-on (SSO) and multi-factor authentication (MFA) to ensure data security.

Images