A channel weight allocation method and device for port machinery equipment status monitoring

Abstract

The present application discloses a method and device for allocating channel weights for monitoring the status of port machinery equipment. The method includes: multiple sensors transmit vibration signal data via Modbus and CAN protocols, and perform FFT transformation on the data to generate an energy spectrum distribution; based on the distribution, wavelet packet decomposition is used to determine the main characteristic frequency band, and its association with the neural network channel is established; based on this association, a multi-layer perceptron is used to perform nonlinear mapping of the frequency band energy, and a dynamic channel weight is generated via a Sigmoid function; the weight is combined with the time series features extracted by Bi‑LSTM for channel weighting, and equipment fault diagnosis is achieved based on the weighted fusion features. The present application extracts key characteristic frequency bands through FFT transformation and wavelet packet decomposition, establishes its association with the neural network channel, and uses a multi-layer perceptron to generate dynamic weights to optimize feature fusion, making equipment fault diagnosis more accurate and improving monitoring efficiency and accuracy.

Description

Technical Field

[0001] The present application relates to the field of port machinery monitoring, and in particular to a channel weight distribution method and device for port machinery equipment status monitoring. Background Art

[0002] Existing port machinery equipment condition monitoring and early warning systems typically consist of a sensor module, a data acquisition and preprocessing module, an intelligent analysis module, and an early warning module. The sensor module is responsible for collecting real-time equipment vibration signals, temperature data, motor current, operating speed, and structural stress and strain data, providing raw information for subsequent analysis. However, in practice, port machinery equipment sensor modules typically use different communication protocols to transmit data, such as Modbus and CAN. These protocols differ in data format, transmission method, and addressing mechanism, making data fusion difficult and affecting the accuracy of equipment condition monitoring.

[0003] Furthermore, regarding data processing, vibration signals often contain significant amounts of noise and irrelevant frequency components. Directly analyzing the raw signal can compromise the accuracy of equipment condition monitoring. Therefore, noise reduction is necessary to extract key features that reflect the health of the equipment. Traditional signal processing methods may not effectively remove noise while retaining key features, resulting in low-quality input data for subsequent intelligent analysis modules, impacting the accuracy of fault diagnosis.

[0004] In intelligent analysis modules, existing technologies typically use a fusion model of a dilated convolutional neural network (DCNN) and a long short-term memory network (LSTM) to extract spatial and time series features. However, traditional fusion methods suffer from issues such as feature redundancy, poor gradient propagation, and wasted computing resources. Specifically, DCNN excels at extracting spatial features, while LSTM excels at capturing long-term dependencies in time series. However, directly fusing these two features can lead to information overlap, compromising model efficiency. Furthermore, the differences in the parameter update mechanisms of the two models in deep networks can cause gradients to vanish or explode, increasing the difficulty of model training. Furthermore, unoptimized fusion structures increase model complexity, making it difficult to meet the low-latency requirements of real-time monitoring and resulting in wasted computing resources. Summary of the Invention

[0005] The purpose of this application is to overcome the defects in the above-mentioned prior art and provide a channel weight allocation method and device for port machinery equipment status monitoring.

[0006] This application provides a channel weight allocation method for port machinery equipment status monitoring, including:

[0007] Acquire vibration signal data transmitted by multiple sensors of port machinery equipment via Modbus and CAN protocols;

[0008] Performing FFT transformation on the vibration signal data to generate energy spectrum distribution;

[0009] Based on the energy spectrum distribution, determining the main characteristic frequency bands through wavelet packet decomposition, and establishing a correlation relationship between the characteristic frequency bands and neural network channels;

[0010] According to the association relationship, a multi-layer perceptron is used to perform nonlinear mapping on the frequency band energy, and a dynamic channel weight is generated through a Sigmoid function;

[0011] The dynamic channel weights are combined with the time series features extracted by Bi-LSTM to perform channel weighting, and equipment fault diagnosis is achieved based on the weighted fusion features.

[0012] Optionally, vibration signal data transmitted by multiple sensors of port machinery equipment via Modbus protocol and CAN protocol is obtained, including:

[0013] The vibration signal data transmitted by the Modbus protocol and the CAN protocol are aligned in the time domain through the precision clock synchronization protocol to eliminate the difference in protocol transmission delay.

[0014] Optionally, based on the energy spectrum distribution, determining a main characteristic frequency band by wavelet packet decomposition, and establishing an association relationship between the characteristic frequency band and a neural network channel include:

[0015] Based on the vibration signal shock attenuation coefficient extracted by wavelet packet decomposition, the frequency band weight distribution ratio in the association relationship is dynamically adjusted.

[0016] Optionally, based on the association relationship, a multilayer perceptron is used to perform nonlinear mapping on the frequency band energy, and a dynamic channel weight is generated by a Sigmoid function, including:

[0017] A channel pruning operation is performed on the dynamic channel weights to remove redundant channels whose weight values ​​are lower than a preset threshold.

[0018] Optionally, the dynamic channel weights are combined with the time series features extracted by the Bi-LSTM to perform channel weighting, and device fault diagnosis is performed based on the weighted fusion features, including:

[0019] The weighted fusion feature hash value is written into the blockchain to achieve tamper-proof storage of equipment fault diagnosis results.

[0020] The present application also provides a channel weight distribution device for port machinery equipment status monitoring, comprising:

[0021] Acquisition module, which acquires vibration signal data transmitted by multiple sensors of port machinery equipment through Modbus protocol and CAN protocol;

[0022] A transformation module performs FFT transformation on the vibration signal data to generate an energy spectrum distribution;

[0023] A relationship module, based on the energy spectrum distribution, determines the main characteristic frequency band through wavelet packet decomposition, and establishes an association relationship between the characteristic frequency band and the neural network channel;

[0024] A weight module, based on the association relationship, uses a multi-layer perceptron to perform nonlinear mapping on the frequency band energy and generates dynamic channel weights through a Sigmoid function;

[0025] The diagnosis module performs channel weighting on the dynamic channel weights and the time series features extracted by the Bi-LSTM, and implements equipment fault diagnosis based on the weighted fusion features.

[0026] Optionally, the acquisition module acquires vibration signal data transmitted by multiple sensors of the port machinery equipment through the Modbus protocol and the CAN protocol, including:

[0027] The vibration signal data transmitted by the Modbus protocol and the CAN protocol are aligned in the time domain through the precision clock synchronization protocol to eliminate the difference in protocol transmission delay.

[0028] Optionally, the relationship module determines a main characteristic frequency band through wavelet packet decomposition based on the energy spectrum distribution, and establishes an association relationship between the characteristic frequency band and the neural network channel, including:

[0029] Based on the vibration signal shock attenuation coefficient extracted by wavelet packet decomposition, the frequency band weight distribution ratio in the association relationship is dynamically adjusted.

[0030] Optionally, the weight module uses a multi-layer perceptron to perform nonlinear mapping on the frequency band energy according to the association relationship, and generates dynamic channel weights through a Sigmoid function, including:

[0031] A channel pruning operation is performed on the dynamic channel weights to remove redundant channels whose weight values ​​are lower than a preset threshold.

[0032] Optionally, the diagnosis module performs channel weighting on the dynamic channel weight and the time series features extracted by the Bi-LSTM, and implements device fault diagnosis based on the weighted fusion features, including:

[0033] The weighted fusion feature hash value is written into the blockchain to achieve tamper-proof storage of equipment fault diagnosis results.

[0034] The beneficial effects of this application are:

[0035] The present application provides a channel weight allocation method for port machinery equipment status monitoring, including: obtaining vibration signal data transmitted by multiple sensors of port machinery equipment through Modbus protocol and CAN protocol; performing FFT transformation on the vibration signal data to generate energy spectrum distribution; based on the energy spectrum distribution, determining the main characteristic frequency band through wavelet packet decomposition, and establishing an association relationship between the characteristic frequency band and the neural network channel; according to the association relationship, using a multi-layer perceptron to perform nonlinear mapping on the frequency band energy, and generating dynamic channel weights through a Sigmoid function; channel-weighting the dynamic channel weights and the time series features extracted by Bi-LSTM, and realizing equipment fault diagnosis based on the weighted fusion features. The present application extracts key characteristic frequency bands through FFT transformation and wavelet packet decomposition, establishes its association with the neural network channel, uses a multi-layer perceptron to generate dynamic channel weights, optimizes feature fusion, realizes more accurate equipment fault diagnosis, and improves monitoring efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a schematic diagram of channel weight distribution for port machinery equipment status monitoring in this application;

[0037] Figure 2 This is a schematic diagram of the structure of the dynamic feature fusion submodule of the intelligent analysis module;

[0038] Figure 3 This is the workflow diagram of the early warning module in this application. DETAILED DESCRIPTION

[0039] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it is understood that various forms of implementing the present disclosure are not limited by the embodiments set forth herein. Rather, the embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0040] Please refer to Figures 1-3 As shown, the present application provides a channel weight allocation method for port machinery equipment status monitoring, comprising:

[0041] S101, obtaining vibration signal data transmitted by multiple sensors of port machinery equipment through Modbus protocol and CAN protocol;

[0042] Multi-source data fusion unit: Sensor modules of large-scale port machinery and equipment usually use different communication protocols to transmit data, including industrial protocols such as Modbus and CAN.

[0043] Modbus is a common serial communication protocol, often used for data exchange between PLCs, sensors, and actuators. CAN bus, on the other hand, is a highly reliable serial communication protocol widely used in vehicles, industrial automation, and embedded systems. Due to differences in data formats, transmission methods, and addressing mechanisms between these protocols, a protocol parser is required to convert data to ensure uniform processing of data from all sensors.

[0044] In the Modbus protocol, sensor data is transmitted through a master-slave architecture. The master sends a request, and the slave returns data. For example, a temperature sensor on a device transmits data via Modbus. After the host sends a read command, the device returns a hexadecimal data frame containing the device address, function code, data length, and actual measurement value. After receiving this data frame, the protocol parser parses the fields according to the Modbus protocol, extracts the temperature value, and converts it into a standard format, such as a floating-point number expressed in degrees Celsius.

[0045] Under the CAN bus protocol, data is transmitted via a broadcast mechanism between nodes. Each data frame contains an identifier, control information, and a data field. For example, if a vibration sensor transmits acceleration values ​​via the CAN bus, its data frame might include a unique identifier and an eight-byte data field. Upon receiving the data, the protocol parser determines the data type based on the identifier and, using predefined parsing rules, extracts the valid information and converts it into a standardized format, such as acceleration values ​​in standard units.

[0046] Through the protocol parser, data from different protocols is converted into consistent data structures, such as timestamps, sensor types, measurement values, and units, enabling subsequent data acquisition and preprocessing modules to efficiently perform data fusion and feature extraction, providing standardized input for the intelligent analysis module and ensuring the accuracy and reliability of the equipment status monitoring and early warning system.

[0047] S102, performing FFT transformation on the vibration signal data to generate an energy spectrum distribution;

[0048] Feature Engineering Module: Extracts important features that reflect the health status of the equipment from the original vibration signal, helping the subsequent intelligent analysis module to identify and predict faults.

[0049] Kurtosis and skewness are commonly used statistical measures for time-domain feature extraction, effectively reflecting signal nonstationarity and sharpness. Kurtosis, the "steepness" of a signal distribution, can capture abnormal fluctuations in the signal and is often used to detect mechanical faults, such as bearing damage. Skewness measures the asymmetry of the signal distribution and can reflect deviations from normal conditions in vibration signals.

[0050] These time-domain features, derived by statistically analyzing transient signal changes, can help identify potential faults or abnormal conditions. By calculating 12-dimensional time-domain features such as kurtosis and skewness, a comprehensive time-domain description of the device's operating status can be provided.

[0051] In terms of frequency domain feature extraction, the energy spectrum peak features after FFT transformation provide detailed information about the signal's frequency components. The FFT transform converts the time domain signal into a frequency domain signal, revealing the energy distribution of different frequency components. By analyzing the frequency domain signal and extracting 10 energy spectrum peak features, it is possible to reveal the concentration of energy in different frequency bands. For example, when a device fails, it may produce an abnormal surge in frequency components within a specific frequency range. Identifying these energy spectrum peaks can help locate the type and location of the fault. For example, a bearing fault may exhibit a significant energy peak within a specific frequency bandwidth.

[0052] By combining time domain features with frequency domain features, the feature engineering module provides a set of multidimensional features containing equipment operating status information. These features can provide an important basis for subsequent intelligent analysis modules, helping to achieve more accurate fault diagnosis and remaining life prediction.

[0053] S103, based on the energy spectrum distribution, determining the main characteristic frequency band through wavelet packet decomposition, and establishing an association relationship between the characteristic frequency band and the neural network channel;

[0054] Signal Noise Reduction Subunit: Vibration signals often contain significant amounts of noise and irrelevant frequency components. Direct analysis of the raw signal can affect the accuracy of equipment condition monitoring. Therefore, the data acquisition and preprocessing module uses the DB4 wavelet packet basis function to perform a three-layer decomposition of the vibration signal to extract key frequency features and remove irrelevant information.

[0055] The db4 wavelet packet basis function belongs to the Daubechies wavelet family, has good compact support and smoothness, and is suitable for processing non-stationary signals.

[0056] During the three-layer wavelet packet decomposition process, the vibration signal is first decomposed using the DB4 basis function, yielding two sub-signals: low-frequency and high-frequency. Each sub-signal is then further decomposed to yield finer-grained frequency components. After three layers of decomposition, the original signal is divided into multiple frequency bands, each containing vibration information from a different frequency range. By calculating the energy contribution of each frequency band, it is possible to determine which bands contain the majority of the signal energy and which contribute less.

[0057] During signal reconstruction, frequency bands with an energy share exceeding 95% are selected for reconstruction. This preserves the key characteristic information while filtering out low-energy noise bands. For example, after wavelet packet decomposition of the bearing vibration signal of a port crane, it was found that certain high-frequency components had a very low energy share, indicating that these components were primarily noise. However, certain components in the mid- and low-frequency bands had higher energy, indicating that they contained critical information about the equipment's operating status. Therefore, during the reconstruction process, low-energy frequency bands are removed, retaining only high-energy signal components. This ensures that the reconstructed signal contains both key vibration information and effectively reduces noise interference, providing more reliable data input for subsequent feature extraction and intelligent analysis.

[0058] S104, performing nonlinear mapping on the frequency band energy using a multilayer perceptron according to the association relationship, and generating dynamic channel weights using a Sigmoid function;

[0059] Channel attention calculation: assign different importance to features of different channels to enhance the focus on key features.

[0060] First, the input feature tensor Perform ReLU activation to ensure that all eigenvalues ​​are non-negative to highlight valid information.

[0061] Then, global average pooling is performed along the time dimension, that is, the average value of each channel over all time steps is calculated to obtain channel statistics:

[0062]

[0063] Among them, Ft represents the value of the input feature at time step t, It represents the average activation value of each channel, reflecting the global contribution of the channel in the entire time series.

[0064] Next, a multi-layer perceptron (MLP) is used to perform feature mapping on Fc to learn the nonlinear relationship between channels and obtain channel weights:

[0065]

[0066] in, represents the channel weight, and σ is the Sigmoid activation function, which ensures that the weight value is between 0 and 1 and is used to measure the importance of each channel. Finally, Wc is weighted with the original features according to the channel dimension, so that high-weight channels contribute more to subsequent calculations, thereby improving the model's discriminative ability.

[0067] Temporal Attention Computation: Dynamically adjust feature weights at different time steps to highlight the impact of key moments.

[0068] Input feature tensor Contains information for all time steps, where B is the batch size, T is the number of time steps, and C is the number of channels.

[0069] In order to extract the importance of the time dimension, we first apply a one-dimensional convolutional layer (Conv1D) to perform a convolution operation along the time dimension to obtain a time-related feature representation:

[0070]

[0071] in, To represent time-dependent features, the convolution kernel of Conv1D slides on the time axis to extract local temporal patterns and capture short-term temporal dependencies. Next, the convolution output is normalized by Softmax to obtain the normalized weight of the time step:

[0072]

[0073] in, Indicates the importance of each time step. Softmax ensures that the sum of all time weights is 1, giving the overall distribution a probabilistic interpretation. Finally, by weighting the original features with Wt, the model pays more attention to key time steps during the decision-making process and improves its sensitivity to changes in device status.

[0074] S105 , performing channel weighting on the dynamic channel weight and the time series features extracted by Bi-LSTM, and implementing device fault diagnosis based on the weighted fusion features.

[0075] Feature fusion: Establish a dynamic balance between spatial features and temporal features so that the model can effectively utilize information from both.

[0076] The spatial features extracted by DCNN are recorded as , the temporal features extracted by Bi-LSTM are recorded as .

[0077] First, through the temporal attention weight Adjust the contribution of the two features and calculate the time-weighted fusion features:

[0078]

[0079] in, is the time-weighted fusion feature, ⊙ represents element-by-element multiplication, and Wt controls the relative contribution of DCNN and Bi-LSTM at each time step. When Wt at a time step is larger, the DCNN feature has a higher weight at that time step, and vice versa.

[0080] Then, using the channel attention weight Further weight the fusion features:

[0081]

[0082] in, To further weight the fusion features, Wc controls the importance of different channels, so that the model can highlight the key channel information in the fusion features. It not only contains the complementary information of DCNN and Bi-LSTM, but also enhances the influence of key time steps and important channels through the attention mechanism, improves the feature expression ability, and provides high-quality input for subsequent fault identification and life prediction.

[0083] Furthermore, the intelligent analysis module further includes:

[0084] Residual connection structure: alleviates the gradient vanishing problem in deep model training while retaining the original input feature information and enhancing feature expression capabilities.

[0085] Fusion Features It has undergone multi-layer feature extraction and attention weighting, and contains comprehensive information of spatial features and temporal features.

[0086] However, in the deep learning process, features may lose some of their original information after multiple transformations, making it difficult for the model to maintain sensitivity to the initial state.

[0087] Therefore, based on the fusion features, the residual connection is introduced to convert the original input features Directly superimpose on the fusion features and calculate the final output:

[0088]

[0089] in, It is the feature after the original input feature and the fusion feature are superimposed.

[0090] This structure ensures that even if the nonlinear transformation in the subsequent network causes information loss, the original input features can still be partially retained, providing additional reference information for the model. At the same time, the residual connection helps the gradient propagate forward more stably, reduces the difficulty of deep network training, and improves the convergence speed and optimization effect. It not only integrates the key features of different channels and time steps, but also maintains the original description of the equipment's operating status, improving the accuracy of the intelligent analysis module in identifying equipment fault types and predicting remaining life.

[0091] In edge computing environments, due to limited computing resources, the lightweight optimization layer needs to improve inference efficiency by optimizing the model structure and reducing computational complexity. Channel pruning technology is based on L1 regularization, which guides the model to automatically identify and remove channels with low contribution, thereby reducing redundant calculations and improving computational efficiency. , add L1 regularization term during training:

[0092]

[0093] in, To add the weight tensor after L1 regularization, is the number of output channels, represents the weight of the i-th channel, λ is the regularization coefficient, is the input channel data, and K is the convolution kernel size. L1 regularization forces some channel weights to gradually approach zero, allowing these channels to be safely removed during pruning, reducing computational overhead while minimizing the impact on model accuracy.

[0094] The pruned model still requires further optimization to accommodate inference on low-computing devices. The INT8 quantized inference module reduces storage requirements and accelerates inference computations by converting model weights and activation values ​​from 32-bit floating-point numbers to 8-bit integers. The quantization process involves scaling weights and activation values, mapping continuous values ​​to discrete integer space while maintaining numerical stability. Using a quantized model can effectively reduce computational load and increase inference speed, enabling the full-state intelligent monitoring and early warning system for large-scale port machinery to operate efficiently in an edge computing environment.

[0095] Furthermore, the dynamic feature fusion submodule of the intelligent analysis module is specifically configured as follows:

[0096] Dual-channel feature alignment:

[0097] Dual-channel feature alignment aims to solve the mismatch problem between DCNN and Bi-LSTM in feature dimensions so as to facilitate subsequent effective fusion.

[0098] DCNN is responsible for extracting spatial features, and the dimension of the output tensor is B×T×C1, where B represents the batch size, T represents the time step, and C1 represents the number of spatial feature channels. Bi-LSTM is used to capture temporal dependencies, and the dimension of the output tensor is B×T×C2, where C2 is the number of temporal feature channels. Since C1 and C2 are usually different, direct splicing or weighted averaging will lead to unbalanced feature information and affect the fusion effect. Therefore, a fully connected layer is used to map the two features to a unified dimension B×T×C respectively to ensure alignment of the information in the channel dimension. Let W1 and W2 be weight matrices, b1 and b2 be bias terms, and the feature mapping process can be expressed as

[0099]

[0100]

[0101] Among them, Fs and Ft represent the original outputs of DCNN and Bi-LSTM respectively, and is the mapped feature, and σ is the activation function, which enhances the nonlinear expression capability. Through this mapping operation, the spatial features and temporal features remain consistent in the channel dimension, laying the foundation for subsequent adaptive attention fusion.

[0102] The early warning module is configured with multi-level threshold settings. By setting different thresholds, it can judge the abnormal status of the equipment and issue early warning signals in time.

[0103] A Level 1 alert is triggered when a device's characteristic value exceeds 1.5 times its historical average, typically indicating a deviation from the device's operating status. At this point, the system issues a yellow alert, alerting personnel to the device's status, warning of potential incipient failure or performance degradation, requiring inspection and monitoring. While yellow alerts generally don't require immediate intervention, close monitoring of the device is required to respond if the problem develops further.

[0104] The triggering conditions for the second-level warning are more stringent. When the characteristic value of the equipment exceeds twice the historical average and lasts for 5 minutes, the system will issue a red warning. This warning usually means that the equipment has experienced a serious abnormality, which may cause equipment failure or damage. Abnormal signals that last for more than 5 minutes indicate that the problem is already urgent and the system needs to take quick measures to prevent the fault from worsening. When a red warning is issued, the system automatically cuts off the control power supply of the equipment to prevent the fault from further expanding or causing greater losses. At the same time, the system will also send an alarm text message containing the fault code to the port dispatch center to ensure that relevant personnel can obtain equipment failure information in a timely manner and take countermeasures.

[0105] For example, if the system detects a vibration characteristic value exceeding 1.5 times the historical average in the vibration signal of a port crane, it will trigger a yellow alert, alerting personnel to pay attention. If the vibration characteristic value continues to rise and exceeds 2 times the historical average within 5 minutes, the system will automatically trigger a red alert, cutting off power and sending an alarm to the dispatch center to prevent further damage caused by excessive equipment wear or failure. This multi-level early warning strategy can effectively improve the accuracy and response speed of fault diagnosis, ensuring the safe operation of equipment.

[0106] Furthermore, the data storage and management module is responsible for the efficient storage, management, and protection of system data, ensuring the security, integrity, and long-term traceability of device monitoring data. This module utilizes a distributed storage architecture, establishing clustered storage nodes based on HDFS (Hadoop Distributed File System). HDFS is highly fault-tolerant, scalable, and cost-effective, effectively supporting the storage and rapid access of large-scale data. The storage period for raw data is set to one year, meaning that real-time collected information such as vibration signals and temperature data from the equipment is retained for one year for short-term analysis and fault tracing. The storage period for analysis results is set to five years, ensuring that historical analysis results from the equipment over a long period of time can provide a reference for subsequent equipment maintenance and optimization decisions. This storage period is designed to balance the use of storage resources with the long-term value of the data.

[0107] In the blockchain evidence storage module, to ensure data immutability and security, the system stores the hash value of key feature vectors on-chain. A hash value is a unique representation of feature data processed by a hash algorithm. By hashing the data, a fixed-length digital fingerprint is generated, ensuring data integrity. Once the hash value of the feature vector is stored on-chain, any subsequent tampering with the data will result in a change in the hash value. This allows verification of data authenticity by comparing the hash value on-chain with the hash value of the stored data. Blockchain consensus mechanisms, such as PBFT (Practical Byzantine Fault Tolerance), ensure data consistency and immutability in a distributed environment through voting and consistency checks between nodes, preventing the risk of data tampering, deletion, or forgery.

[0108] For example, when a port crane malfunctions, the system records and analyzes the relevant characteristic data to generate fault diagnosis results. These analysis results are stored using blockchain technology. This ensures that even if the equipment's monitoring data is accessed or modified by external systems, historical fault diagnosis results can still be verified through the blockchain, maintaining data transparency and reliability. This blockchain-based evidence storage mechanism provides credible digital evidence of the equipment's operating history, helping to improve the transparency of equipment maintenance and ensure compliance with operational regulations.

[0109] The present application also provides a channel weight distribution device for port machinery equipment status monitoring, comprising:

[0110] Acquisition module, which acquires vibration signal data transmitted by multiple sensors of port machinery equipment through Modbus protocol and CAN protocol;

[0111] A transformation module performs FFT transformation on the vibration signal data to generate an energy spectrum distribution;

[0112] A relationship module, based on the energy spectrum distribution, determines the main characteristic frequency band through wavelet packet decomposition, and establishes an association relationship between the characteristic frequency band and the neural network channel;

[0113] A weight module, based on the association relationship, uses a multi-layer perceptron to perform nonlinear mapping on the frequency band energy and generates dynamic channel weights through a Sigmoid function;

[0114] The diagnosis module performs channel weighting on the dynamic channel weights and the time series features extracted by the Bi-LSTM, and implements equipment fault diagnosis based on the weighted fusion features.

[0115] Furthermore, the acquisition module acquires vibration signal data transmitted by multiple sensors of the port machinery equipment through the Modbus protocol and the CAN protocol, including:

[0116] The vibration signal data transmitted by the Modbus protocol and the CAN protocol are aligned in the time domain through the precision clock synchronization protocol to eliminate the difference in protocol transmission delay.

[0117] Furthermore, the relationship module determines the main characteristic frequency bands through wavelet packet decomposition based on the energy spectrum distribution, and establishes an association relationship between the characteristic frequency bands and the neural network channels, including:

[0118] Based on the vibration signal shock attenuation coefficient extracted by wavelet packet decomposition, the frequency band weight distribution ratio in the association relationship is dynamically adjusted.

[0119] Furthermore, the weight module uses a multi-layer perceptron to perform nonlinear mapping on the frequency band energy according to the association relationship, and generates dynamic channel weights through a Sigmoid function, including:

[0120] A channel pruning operation is performed on the dynamic channel weights to remove redundant channels whose weight values ​​are lower than a preset threshold.

[0121] Furthermore, the diagnosis module performs channel weighting on the dynamic channel weights and the time series features extracted by the Bi-LSTM, and implements device fault diagnosis based on the weighted fusion features, including:

[0122] The weighted fusion feature hash value is written into the blockchain to achieve tamper-proof storage of equipment fault diagnosis results.

[0123] The above description of the embodiments is intended to facilitate understanding and application by those skilled in the art. Those skilled in the art will readily be able to make various modifications to the above embodiments and apply the general principles described herein to other embodiments without requiring inventive effort. Therefore, without being limited to the above embodiments, improvements and modifications made by those skilled in the art based on the disclosure of the present invention should be within the scope of protection of the present invention.

Claims

1. A channel weight allocation method for port machinery equipment status monitoring, characterized in that: include: Acquire vibration signal data transmitted by multiple sensors of port machinery equipment via Modbus and CAN protocols; Performing FFT transformation on the vibration signal data to generate energy spectrum distribution; Based on the energy spectrum distribution, determining characteristic frequency bands through wavelet packet decomposition, and establishing an association relationship between the characteristic frequency bands and neural network channels; According to the association relationship, a multi-layer perceptron is used to perform nonlinear mapping on the energy of the characteristic frequency band, and a dynamic channel weight is generated through a Sigmoid function; The dynamic channel weights are combined with the time series features extracted by Bi-LSTM to perform channel weighting, and equipment fault diagnosis is achieved based on the weighted fusion features.

2. A channel weight allocation method for port machinery equipment status monitoring according to claim 1, characterized in that: Acquire vibration signal data transmitted by multiple sensors of port machinery equipment via Modbus and CAN protocols, including: The vibration signal data transmitted by the Modbus protocol and the CAN protocol are aligned in the time domain through the precision clock synchronization protocol to eliminate the difference in protocol transmission delay.

3. The channel weight allocation method for port machinery equipment status monitoring according to claim 1 is characterized in that: Based on the energy spectrum distribution, characteristic frequency bands are determined by wavelet packet decomposition, and an association relationship between the characteristic frequency bands and neural network channels is established, including: Based on the impact attenuation coefficient of the vibration signal extracted by wavelet packet decomposition, the frequency band weight distribution ratio in the association relationship is dynamically adjusted.

4. The channel weight allocation method for port machinery equipment status monitoring according to claim 1 is characterized in that: According to the association relationship, a multi-layer perceptron is used to perform nonlinear mapping on the characteristic frequency bands, and a dynamic channel weight is generated by a Sigmoid function, including: A channel pruning operation is performed on the dynamic channel weights to remove redundant channels whose weight values ​​are lower than a preset threshold.

5. The channel weight allocation method for port machinery equipment status monitoring according to claim 1 is characterized in that: The dynamic channel weights are combined with the time series features extracted by the Bi-LSTM to perform channel weighting, and device fault diagnosis is performed based on the weighted fusion features, including: The hash value of the weighted fusion feature is written into the blockchain to achieve tamper-proof storage of the equipment fault diagnosis result.

6. A channel weight distribution device for port machinery equipment status monitoring, characterized in that: include: Acquisition module, which acquires vibration signal data transmitted by multiple sensors of port machinery equipment through Modbus protocol and CAN protocol; A transformation module performs FFT transformation on the vibration signal data to generate an energy spectrum distribution; a relationship module, which determines characteristic frequency bands by wavelet packet decomposition based on the energy spectrum distribution, and establishes an association relationship between the characteristic frequency bands and neural network channels; The weight module uses a multi-layer perceptron to perform nonlinear mapping on the energy of the characteristic frequency band according to the association relationship, and generates dynamic channel weights through a Sigmoid function; The diagnosis module performs channel weighting on the dynamic channel weights and the time series features extracted by the Bi-LSTM, and implements equipment fault diagnosis based on the weighted fusion features.

7. A channel weight distribution device for port machinery equipment status monitoring according to claim 6, characterized in that: The acquisition module acquires vibration signal data transmitted by multiple sensors of port machinery equipment through Modbus protocol and CAN protocol, including: The vibration signal data transmitted by the Modbus protocol and the CAN protocol are aligned in the time domain through the precision clock synchronization protocol to eliminate the difference in protocol transmission delay.

8. The channel weight distribution device for port machinery equipment status monitoring according to claim 6 is characterized in that: The relationship module determines a characteristic frequency band based on the energy spectrum distribution by wavelet packet decomposition, and establishes an association relationship between the characteristic frequency band and the neural network channel, including: Based on the impact attenuation coefficient of the vibration signal extracted by wavelet packet decomposition, the frequency band weight distribution ratio in the association relationship is dynamically adjusted.

9. The channel weight distribution device for port machinery equipment status monitoring according to claim 6 is characterized in that: The weight module uses a multi-layer perceptron to perform nonlinear mapping on the characteristic frequency band according to the association relationship, and generates dynamic channel weights through a Sigmoid function, including: A channel pruning operation is performed on the dynamic channel weights to remove redundant channels whose weight values ​​are lower than a preset threshold.

10. The channel weight distribution device for port machinery equipment status monitoring according to claim 6, characterized in that: The diagnosis module performs channel weighting on the dynamic channel weights and the time series features extracted by the Bi-LSTM, and implements device fault diagnosis based on the weighted fusion features, including: The hash value of the weighted fusion feature is written into the blockchain to achieve tamper-proof storage of the equipment fault diagnosis result.

Images