A crane reduction gearbox state monitoring method based on multivariable proportional regulation

Abstract

The application discloses a crane reduction gearbox state monitoring method based on multivariate proportional regulation, comprising collecting two-dimensional operation data of the crane reduction gearbox; acquiring working condition information of crane operation through a motor frequency converter; inputting the obtained data into a neural network model, adopting multivariate proportional regulation for multivariate linear regression model fitting; extracting features of the obtained data by using a global multilevel channel attention mechanism, completing data decision, minimizing a loss function through a back propagation algorithm, and using an optimization algorithm to accelerate the learning process; inputting the data into a deep learning model to complete crane reduction gearbox health state monitoring, and verifying the accuracy of the deep learning model.

Description

Technical Field

[0001] This invention relates to the field of condition monitoring technology, and in particular to a method for condition monitoring of a multi-proportional adjustment deep learning gearbox that combines operating condition data. Background Technology

[0002] As the workload of large port machinery becomes increasingly heavy, enterprises have higher and higher requirements for the management efficiency and safety performance of large port machinery and equipment. Therefore, how to utilize advanced information technology to improve management efficiency and enhance the safety level of port machinery is an important issue facing port enterprises. Gantry cranes, with their advantages of high operating efficiency, high degree of automation, and low operating costs, have been favored by major ports. In recent years, factors such as government regulation, industry competition, and technological advancements have driven new business models and trends in the automation and unmanned operation of gantry cranes.

[0003] The automation transformation of gantry cranes has become a major development direction for ports, with the upgrading and transformation of automated equipment gradually underway. However, the gearbox is a vulnerable component of gantry cranes, and existing monitoring methods cannot detect abnormal conditions during operation. The health status monitoring of gantry crane gearboxes suffers from low efficiency and accuracy, necessitating model improvements. Against this backdrop, relevant research was conducted, addressing the online monitoring needs of gearboxes in the hoisting, closing, slewing, and luffing mechanisms of gantry cranes. By combining real-time operating conditions of the port cranes, a comprehensive real-time status assessment was performed to scientifically guide users in equipment maintenance, thereby reducing equipment failure rates and management costs, and ultimately achieving cost reduction and efficiency improvement. Summary of the Invention

[0004] To overcome the aforementioned problems in the existing technology, this invention proposes a crane gearbox condition monitoring method based on multivariable proportional adjustment.

[0005] The technical solution adopted by this invention to solve its technical problem is: a method for monitoring the condition of a crane gearbox based on multivariable proportional adjustment, comprising the following steps:

[0006] Step 1: Collect two-dimensional operating data of the crane's gearbox;

[0007] Step 2: Obtain the crane's operating condition information through the motor frequency converter;

[0008] Step 3: Input the data obtained in Step 1 and Step 2 into the neural network model, and use multivariate proportional adjustment to fit the multiple linear regression model;

[0009] Step 4: Extract features from the data obtained in Step 3 using a global multi-level channel attention mechanism to complete data decision-making. Minimize the loss function using the backpropagation algorithm and use an optimization algorithm to accelerate the learning process.

[0010] Step 5: Input the data from Step 4 into the deep learning model to complete the health status monitoring of the crane gearbox and verify the accuracy of the deep learning model.

[0011] The above-mentioned method for monitoring the condition of a crane gearbox based on multivariable proportional control, specifically includes step 3 as follows:

[0012] Step 3.1, Data Preprocessing: Data cleaning and outlier calculation are performed on the data obtained in Steps 1 and 2. Outliers are removed from the samples through a traversal process. The specific formulas for data cleaning and outlier calculation are as follows:

[0013] Y = (x i -μ) / σ

[0014] Where Y represents the deviation value used to identify outliers, x represents the value of the sample point, i represents the sample number, μ represents the mean of the sample population, and σ represents the standard deviation of the sample population;

[0015] Step 3.2: Input the data obtained in Step 3.1 into the neural network model to obtain the global feature vector;

[0016] Step 3.3: Perform correlation and causality analysis on the data obtained in Step 3.1 to determine the impact of each factor on the performance of the crane;

[0017] Step 3.4: Based on the analysis results of Step 3.3, an attention mechanism is applied to the global feature vector obtained in Step 3.2 to adjust the proportions of different variables in the data obtained in Step 3.1 and obtain the fitted data.

[0018] The above-mentioned method for monitoring the condition of a crane gearbox based on multivariable proportional adjustment, wherein the formula for calculating the fitted data in step 3.4 is as follows:

[0019]

[0020] w α =sign(r) α )*log(n α / (1-n α )) / log(k α )

[0021] w β =sign(r) β )*log(n β / (1-n β )) / log(k β )

[0022] w γ =sign(r) γ )*log(nγ / (1-n γ )) / log(k γ )

[0023] w θ =sign(r) θ )*log(n θ / (1-n θ )) / log(k θ )

[0024] Among them, W t VT(a,b) represents the fitted data at time point t, α represents the crane current operating coefficient, β represents the crane voltage operating coefficient, γ represents the crane power operating coefficient, θ represents the crane speed operating coefficient, V(t) represents the vibration data at time point t, T(t) represents the temperature data at time point t, ψ is the mother wavelet function, and a and b are the scale and translation parameters, respectively; m is the value of the vibration weight index at time point t, n is the value of each temperature index at time point t, and t represents the data time; w α It is the weight of the α factor, r α It is the relative relationship coefficient of the α factor, n α k is the importance coefficient of the α factor. α It is a constant related to factor α; w β It is the weight of the β factor, r β It is the relative relationship coefficient of the β factor, n β k is the importance coefficient of the β factor. β It is a constant related to factor β; w γ It is the weight of the γ factor, r γ It is the relative relationship coefficient of the γ factor, n γ k is the importance coefficient of the γ factor. γ It is a constant related to the factor γ.

[0025] The above-mentioned method for monitoring the condition of a crane gearbox based on multivariable proportional control, specifically includes step 4 as follows:

[0026] Step 4.1: Extract basic features using standard convolutional layers; use a global multi-level channel attention module to weight the features. This step includes multiple attention layers, each of which processes the output of the previous layer to generate a weight distribution, which is used to weight the original features.

[0027] Step 4.2: Calculate the derivative of the loss function with respect to each parameter in the neural network model, and then update the parameters according to the direction of the derivative to reduce the value of the loss function.

[0028] The specific expression in step 4.1 of the above-mentioned crane gearbox condition monitoring method based on multivariable proportional adjustment is as follows:

[0029] output_vector = [Channel] 1 Channel 2 Channel N ]

[0030] Here, `output_vector` represents a multi-channel vector output vector, which is composed of stacked vectors from multiple channels. 1 Channel 2 ...Channel N Each channel represents a vector. Semicolons are used to separate the vectors for each channel, and they are then concatenated to form a higher-level vector.

[0031] in,

[0032]

[0033] Among them, W t VT(a,b) represents the input fused data; i represents the W... t VT(a,b) represents the number of equal parts, i∈{0,1...,t}; W t VT(a,b)∈R C′*H*W , C represents the dimension divisible by t, H represents the height of the dimension, W represents the width of the dimension, and channel. i ∈R C′ This is the preprocessed C'-dimensional vector.

[0034] In the above-mentioned crane gearbox condition monitoring method based on multivariate proportional adjustment, the optimization algorithm in step 4 is the gradient descent algorithm.

[0035] The above-mentioned method for monitoring the condition of a crane gearbox based on multivariable proportional adjustment includes, in step 5, verifying the model accuracy, which includes:

[0036] Step 5.1: Calculate the model training accuracy value based on the crane gearbox health status classification results. The specific formula for calculating the accuracy value is as follows:

[0037]

[0038] Where TP represents the number of correct results from the deep learning model, FP represents the number of incorrect results from the deep learning model, and P represents the accuracy of the deep learning model results.

[0039] Step 5.2: Plot the loss curve and observe the changes in the model's loss during training. If the loss curve tends to be stable, the model has converged. If the loss curve fluctuates greatly, it indicates that the model has not converged.

[0040] The beneficial effects of this invention are as follows: by combining the operating condition information of the gearbox of the gantry crane to gain a refined understanding of the gearbox's health status under different operating conditions, and by combining multi-dimensional operating data such as temperature and acceleration data, all variables are proportionally adjusted to form a comprehensive index evaluation model. The multi-proportional adjustment method comprehensively compares and analyzes the acquired data to determine influencing factors and then proportionally adjusts different variables accordingly. By using a global multi-level channel attention mechanism to extract features, and combining the real-time operating conditions of the port machinery, a comprehensive real-time status evaluation is conducted to scientifically guide users in equipment maintenance, improve the model's detection accuracy, reduce equipment failure rate and management costs, improve the accuracy of health monitoring, and achieve cost reduction and efficiency improvement. Attached Figure Description

[0041] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0042] Figure 1 This is a flowchart of the present invention;

[0043] Figure 2 This is a structural diagram of the monitoring method of the present invention;

[0044] Figure 3 This is a diagram showing the results of a field data comparison experiment for the present invention.

[0045] Figure 4 This is a diagram showing the experimental test results of the field data of this invention;

[0046] Figure 5 This is a graph showing the comparison results between the open-source dataset of this invention and other models. Detailed Implementation

[0047] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] This embodiment discloses a deep learning-based method for monitoring the health status of a gantry crane gearbox based on multivariate proportional adjustment, such as... Figure 1-2 As shown, it includes the following steps:

[0049] Step 1: Measure the temperature and vibration data of the gantry crane gearbox by deploying an integrated temperature and vibration sensor and a temperature sensor. The sensor is deployed magnetically for easy installation and removal, thereby obtaining two-dimensional operating data of the gantry crane gearbox.

[0050] Step 2: Combine the operating condition data of the gantry crane obtained by the equipment frequency converter, and collect important information on the operation of the gearbox, including current, voltage, speed, etc., to evaluate the operating status of the gantry crane.

[0051] Step 2.1: Measure temperature and vibration to obtain two-dimensional operating data of the crane gearbox. Simultaneously measure and connect the motor frequency converter to collect important information about the gearbox operation, including current, voltage, speed, etc., and use it to evaluate the operating status of the crane.

[0052] Step 3: The comprehensive analysis of the data obtained in Steps 1 and 2 is a complex process, but it is an effective method to determine the impact of various factors on the performance of the crane's gearbox. The gearbox health status is combined with equipment operating information to gain a refined understanding of the gearbox's health status under different operating conditions. Equipment operating conditions, temperature, vibration, and other variables are proportionally adjusted to form a comprehensive index evaluation model. The gearbox data is input into a multivariate proportional adjustment convolutional neural network model, and a multivariate proportional adjustment global multi-level channel attention model is used for fitting, facilitating the evaluation of the crane's performance and status in Step 4.

[0053] Step 3.1, Data Preprocessing: First, the acquired temperature, vibration and operating condition data are preprocessed, including data cleaning and outlier handling, to ensure the accuracy and reliability of the data and provide a solid foundation for subsequent analysis.

[0054] Data cleaning and outlier statistics methods are used to assess the position of a sample point relative to the population. The sample point is subtracted from the mean of the population and then divided by the standard deviation of the population to obtain a standardized score.

[0055] The formula for data cleaning and outlier calculation is: Y = (x i -μ) / σ

[0056] Where Y represents the deviation value used to identify outliers, x represents the value of the sample point, i represents the sample number, μ represents the mean of the sample population, and σ represents the standard deviation of the sample population.

[0057] Remove outliers from the sample by iterating through it.

[0058] data = np.array([...])

[0059] mu = np.mean(data)

[0060] sigma = np.std(data)

[0061] Y=[(x-mu) / sigma for x in data]

[0062] Y=[y for y in Y if not np.isnan(y)]

[0063] Y is used to compare the position of different sample points within the population. A positive Y indicates that the sample point's value is greater than the population mean; a negative Y indicates that the sample point's value is less than the population mean; and a Y of 0 indicates that the sample point's value is equal to the population mean. Furthermore, the larger the absolute value of Y, the greater the difference between the sample point and the population. Sample points with Y values ​​greater than 3 can be considered outliers and can be used to compare differences between different samples, such as comparing the differences between the experimental control group and the experimental group.

[0064] Step 3.2: CNNs typically consist of multiple convolutional and pooling layers to extract features from the data and capture local data details. To obtain overall global features, a global average pooling layer is added to the network. This layer performs average pooling on each feature map to obtain a global feature vector.

[0065] Step 3.2, Correlation and Causality Analysis: Conduct correlation and causal analysis on the operating condition data, temperature data, and acceleration data to determine the interrelationships and influences among them, and thus identify factors that have a direct or indirect impact on the performance of the crane.

[0066] Step 3.3, Factor Analysis and Multi-Proportion Adjustment: Combining the gearbox operating condition information, an attention mechanism is applied to the global feature vector to enable the network to perceive the important features of the gearbox under different operating conditions in a refined manner. Through factor analysis, the main factors or components representing the potential factors affecting the performance of the crane gearbox are extracted from multiple variables, thereby reducing the number of decision variables and adjusting the proportions of different variables.

[0067]

[0068] w α =sign(r) α )*log(n α / (1-n α )) / log(k α )

[0069] w β =sign(r) β )*log(n β / (1-n β )) / log(k β )

[0070] w γ =sign(r) γ )*log(n γ / (1-nγ )) / log(k γ )

[0071] w θ =sign(r) θ )*log(n θ / (1-n θ )) / log(k θ )

[0072] Among them, w α It is the weight of the α factor, r α It is the relative relationship coefficient of the α factor, n α k is the importance coefficient of the α factor. α It is a constant related to factor α.

[0073] relative relationship coefficient r α It is determined based on the correlation between factors. If factor α is strongly positively correlated with other factors, then r α Take 1; if factor α is strongly negatively correlated with other factors, then r α Take -1; if factor α is uncorrelated with other factors, then r α Take 0.

[0074] Importance coefficient n α It is determined based on the decision-maker's level of importance to factor α. If the decision-maker considers factor α to be very important, then n... α Set to 1; if the decision-maker considers factor α to be more important, then n α Take 0.5; if the decision-maker considers factor α to be less important, then n α Take 0.

[0075] constant k α It is a parameter related to factor α, and its specific value needs to be determined based on the actual situation.

[0076] Similarly, the relevant weights of β, γ, and θ can be calculated.

[0077] Among them, W t VT(a,b) represents the fitted data at time point t, where α represents the current operating coefficient of the gantry crane, β represents the voltage operating coefficient, γ represents the power operating coefficient, θ represents the speed operating coefficient, V(t) represents the vibration data at time point t, T(t) represents the temperature data at time point t, ψ is the mother wavelet function, and a and b are the scale and translation parameters, respectively. m is the value of the vibration weight index at time point t, n is the value of each temperature index at time point t, and t represents the data time.

[0078] This formula means that at each time point, vibration and temperature data are multiplied and assigned different weights, and then the results are synthesized into a new data sequence through wavelet transform. The purpose of this is to capture the correlation between the two data sources and fuse them in a way that reveals information hidden in the original data.

[0079] Step 4: Extract features from the data from Step 3 using a global multi-level channel attention mechanism. This mechanism improves the model's recognition performance by weighting the input data across different spatial and channel dimensions.

[0080] Step 4.1, Construct the basic convolutional network: Use standard convolutional layers to extract basic features. These convolutional layers can include common operations such as padding, stride, activation functions, etc. After extracting the basic features, use a global multi-level channel attention module to weight the features. This module typically includes multiple attention layers, each of which processes the output of the previous layer to generate a weight distribution used to weight the original features.

[0081] Step 4.2, the specific expression is as follows:

[0082] output_vector = [Channel] 1 Channel 2 Channel N ]

[0083] Here, `output_vector` represents a multi-channel vector output vector, which is composed of stacked vectors from multiple channels. 1 Channel 2 ...Channel N Each channel represents a vector. Semicolons are used to separate the vectors for each channel, and they are then concatenated to form a higher-level vector.

[0084] in,

[0085]

[0086] Among them, W t VT(a,b) represents the input fused data; i represents the W... t VT(a,b) represents the number of equal parts, i∈{0,1...,t}; W t VT(a,b)∈R C′*H*W , C represents the dimension divisible by t, H represents the height of the dimension, W represents the width of the dimension, and channel. i ∈R C′This is the preprocessed C'-dimensional vector.

[0087] Step 4.3 involves calculating the derivative of the loss function with respect to each parameter in the network, and then updating the parameters based on the direction of the derivative to reduce the value of the loss function. Optimization algorithms, on the other hand, refer to adjusting the parameters in a specific way to converge to the minimum value of the function more quickly. The following are the relevant formulas for the backpropagation algorithm and optimization algorithms:

[0088] The loss function measures the difference between the model's predictions and the actual results. Assume the prediction result is... If the true result is y, then commonly used loss functions include cross-entropy:

[0089]

[0090] The purpose of the backpropagation algorithm is to calculate the gradient of each parameter in order to update it. Assuming the current layer is l and the loss function is J, the gradient can be calculated using the chain rule:

[0091]

[0092] Among them W l These are the parameters of layer l, Z l It is the output of layer l.

[0093] Gradient descent is a common optimization algorithm used to update the parameters in a model based on gradients. Assuming the current parameters are W and the learning rate is η, the parameter update method is as follows:

[0094]

[0095] in This represents the gradient of the loss function J with respect to the parameter W.

[0096] Global pooling is performed on the feature maps of each channel to obtain summary information for each channel. Multi-level feature aggregation is then used to obtain the attention weights for each channel from the pooled features. The original features of each channel are multiplied by their corresponding attention weights to obtain weighted features. After multiple rounds of this global multi-level channel attention module processing, the features are finally mapped to the target class space through fully connected layers or other structures. The backpropagation algorithm continuously adjusts the parameters to minimize the loss function, and optimization algorithms are used to accelerate the learning process, resulting in the final classification result.

[0097] Step 5: Input the data from Step 4 into the deep learning model to complete the health status monitoring of the gantry crane gearbox and verify the model's accuracy. During model training, labeled data can be used for supervised learning. The labeled data includes the health status, fault type, and wear degree of the gantry crane. Through continuous iterative optimization, the model can gradually learn the mapping relationship between input data and output results. Finally, test data is used to verify the model's accuracy.

[0098] Step 5.1: The training accuracy value is calculated based on the health status classification results of the gantry crane gearbox data monitoring. Comparing the training results of different models on the same dataset allows for the evaluation of model performance. The accuracy value is expressed as follows:

[0099]

[0100] Where TP represents the number of correct results from the deep learning model, FP represents the number of incorrect results from the deep learning model, and P represents the accuracy of the deep learning model results.

[0101] Step 5.2 analyzes the model's convergence speed during training to understand its training efficiency. If the model converges quickly in the early stages of training, it indicates that the model has learned the main features of the dataset, resulting in fast convergence. Conversely, if the model converges slowly during training, it may be necessary to adjust hyperparameters or increase the dataset size. The network model's parameters are updated during training, and training stops when the maximum number of iterations is reached or the training accuracy reaches 0.9999.

[0102] Step 5.3: By plotting the loss curve, you can observe the changes in the model's loss during training. If the loss curve tends to be stable, it indicates that the model has converged, and there may be overfitting or underfitting. If the loss curve fluctuates greatly, it indicates that the model needs more training.

[0103] The beneficial effects of this invention are as follows: by combining the operating condition information of the gearbox of the gantry crane to gain a refined understanding of the gearbox's health status under different operating conditions, and by combining multi-dimensional operating data such as temperature and acceleration data, all variables are proportionally adjusted to form a comprehensive index evaluation model. The multi-proportional adjustment method comprehensively compares and analyzes the acquired data to determine influencing factors and then proportionally adjusts different variables accordingly. By using a global multi-level channel attention mechanism to extract features, and combining the real-time operating conditions of the port machinery, a comprehensive real-time status evaluation is conducted to scientifically guide users in equipment maintenance, improve the model's detection accuracy, reduce equipment failure rate and management costs, improve the accuracy of health monitoring, and achieve cost reduction and efficiency improvement.

[0104] Experimental verification:

[0105] This embodiment of the training was implemented using an NVIDIA GeForce RTX 3060 with 16GB of RAM, Windows 11, Tensorflow 1.7 framework, and Python 3.7. Validation was performed using field-collected datasets, the Case Western Reserve University dataset, the MFPT dataset, and the Paderborn University dataset. The field data is used as a comparison of the training results, as shown in the image. Figure 3 The model converges after 100 training epochs, and the results tend to stabilize after 200 epochs, with a training accuracy of 99.98% and a model loss of 0.0021. Compared to other mainstream algorithms, the evaluation metrics of the deep learning model based on multivariate proportional adjustment proposed in this embodiment—accuracy, recall, F1 score, and area under the ROC curve—show significant improvement in performance by incorporating actual equipment operating conditions and proportionally adjusting various variables. Figure 4 The method was validated in practical applications using data from the gearboxes of gantry cranes collected on-site. The accuracy rate reached 96.18%, higher than other models, demonstrating its effectiveness in monitoring the health status of gearboxes. Validation was also performed on publicly available datasets such as the Case Western Reserve University dataset, the MFPT dataset, and the Paderborn University dataset. Figure 5 The model proposed in this paper also performs well, with all indicators exceeding those of other models, demonstrating the superiority of the proposed model.

[0106] Table 1: Comparison of Results for Each Model Across Four Datasets

[0107]

[0108]

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for monitoring the condition of a crane gearbox based on multivariable proportional control, characterized in that, Includes the following steps: Step 1: Collect two-dimensional operating data of the crane's gearbox; Step 2: Obtain the crane's operating condition information through the motor frequency converter; Step 3: Input the data obtained in Step 1 and Step 2 into the neural network model, and use multivariate proportional adjustment to fit the multiple linear regression model; Step 4: Extract features from the data obtained in Step 3 using a global multi-level channel attention mechanism to complete data decision-making. Minimize the loss function using the backpropagation algorithm and use an optimization algorithm to accelerate the learning process. Step 5: Input the data from Step 4 into the deep learning model to complete the health status monitoring of the crane gearbox and verify the accuracy of the deep learning model. Step 3 specifically includes: Step 3.1, Data Preprocessing: Data cleaning and outlier calculation are performed on the data obtained in Steps 1 and 2. Outliers are removed from the samples through a traversal process. The specific formulas for data cleaning and outlier calculation are as follows: Y = (x i - μ) / σ Where Y represents the deviation value used to identify outliers, x represents the value of the sample point, i represents the sample number, μ represents the mean of the sample population, and σ represents the standard deviation of the sample population; Step 3.2: Input the data obtained in Step 3.1 into the neural network model to obtain the global feature vector; Step 3.3: Perform correlation and causality analysis on the data obtained in Step 3.1 to determine the impact of each factor on the performance of the crane; Step 3.4: Based on the analysis results of Step 3.3, an attention mechanism is applied to the global feature vector obtained in Step 3.2 to adjust the proportions of different variables in the data obtained in Step 3.1 and obtain the fitted data. The formula for calculating the fitted data in step 3.4 is as follows: Among them, W t VT(a,b) represents the fitted data at time point t, α represents the crane current operating coefficient, β represents the crane voltage operating coefficient, γ represents the crane power operating coefficient, θ represents the crane speed operating coefficient, V(t) represents the vibration data at time point t, T(t) represents the temperature data at time point t, ψ is the mother wavelet function, and a and b are the scale and translation parameters, respectively; m is the value of the vibration weight index at time point t, n is the value of each temperature index at time point t, and t represents the data time; w α It is the weight of the α factor, r α It is the relative relationship coefficient of the α factor, n α k is the importance coefficient of the α factor. α It is a constant related to factor α; w β It is the weight of the β factor, r β It is the relative relationship coefficient of the β factor, n β k is the importance coefficient of the β factor. β It is a constant related to factor β; w γ It is the weight of the γ factor, r γ It is the relative relationship coefficient of the γ factor, n γ k is the importance coefficient of the γ factor. γ It is a constant related to the factor γ.

2. The method for monitoring the condition of a crane gearbox based on multivariable proportional adjustment according to claim 1, characterized in that, Step 4 specifically includes: Step 4.1: Extract basic features using standard convolutional layers; use a global multi-level channel attention module to weight the features. This step includes multiple attention layers, each of which processes the output of the previous layer to generate a weight distribution, which is used to weight the original features. Step 4.2: Calculate the derivative of the loss function with respect to each parameter in the neural network model, and then update the parameters according to the direction of the derivative to reduce the value of the loss function.

3. The method for monitoring the condition of a crane gearbox based on multivariable proportional adjustment according to claim 2, characterized in that, The specific expression in step 4.1 is as follows: Here, `output_vector` represents a multi-channel output vector, which is composed of stacked vectors from multiple channels; `Channel` 1 Channel 2 ... Channel N Each channel represents a vector; use semicolons to separate the vectors of each channel and concatenate them together to form a higher-level vector; in, Among them, W t VT(a,b) represents the input fused data; i represents the W... t VT(a,b) represents the number of equal parts. ; C represents the dimension divisible by t, H represents the height of the dimension, and W represents the width of the dimension. For preprocessed C ’ Dimensional vector.

4. The method for monitoring the condition of a crane gearbox based on multivariable proportional adjustment according to claim 1, characterized in that, The optimization algorithm in step 4 is the gradient descent algorithm.

5. The method for monitoring the condition of a crane gearbox based on multivariable proportional adjustment according to claim 1, characterized in that, Step 5, the verification of model accuracy, includes: Step 5.1: Calculate the model training accuracy value based on the crane gearbox health status classification results. The specific formula for calculating the accuracy value is as follows: Where TP represents the number of correct results from the deep learning model, FP represents the number of incorrect results from the deep learning model, and P represents the accuracy of the deep learning model results. Step 5.2: Plot the loss curve and observe the changes in the model's loss during training. If the loss curve tends to be stable, the model has converged. If the loss curve fluctuates greatly, it indicates that the model has not converged.

Images