A load analysis method for intelligent winch mechanism
Through multi-scale clustering analysis and deep reinforcement learning technology, the load fluctuation characteristics of the winch mechanism are identified and optimized, and the problem of response lag in traditional load monitoring methods is solved, deeper data analysis and dynamic optimization are achieved, and the stability of the equipment is significantly improved. and energy efficiency.
Patent Information
- Application Number
- CN202411148791.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-21
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-08-21
AI Technical Summary
Traditional load monitoring methods rely on static data analysis and cannot accurately reflect load fluctuations in complex working environments, resulting in lagging equipment adjustment responses, increasing energy waste and failure risks.
Using multi-scale clustering analysis framework and deep reinforcement learning technology, we can collect and analyze the load data of the winch mechanism, identify short-term, medium-term and long-term load fluctuations, build load mode relationship mapping, and deploy adjustment strategies in the actual environment to achieve dynamic Adapt and optimize control.
In-depth analysis and dynamic optimization control of load conditions are achieved, energy waste is reduced, fault prevention capabilities and equipment stability are improved, and long-term reliability of the winch mechanism is enhanced.
Smart Images

Figure CN118656605B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of load monitoring, and in particular to a load analysis method for an intelligent hoisting mechanism. Background Art
[0002] The field of load monitoring technology focuses on real-time monitoring and analysis of the workload of various mechanical or electronic equipment. Load monitoring technology plays a key role in improving equipment efficiency, preventing failures, and extending equipment life. In intelligent winch mechanisms, load monitoring is particularly important because it is directly related to the safe operation and work efficiency of the mechanism. By monitoring the load, it can ensure that the winch mechanism works within a safe load range, prevent overload conditions, and optimize its working performance.
[0003] Among them, the intelligent winch mechanism load analysis method is a method that uses intelligent technology to monitor and analyze the winch mechanism load in real time. Its purpose is to ensure that the winch mechanism can operate in the best condition and avoid equipment damage or accidents caused by improper use such as overloading. By analyzing the load data, this method can monitor the working status of the winch mechanism in real time, discover potential risks and problems in time, and take corresponding preventive or corrective measures to improve safety and efficiency.
[0004] Traditional load monitoring methods rely on static, single-timescale data analysis, lacking in-depth understanding of load fluctuations and real-time dynamic adjustment capabilities. This approach often fails to accurately reflect the actual load conditions in complex working environments, resulting in delayed response to equipment adjustments, increased energy waste and the risk of failure. For example, when encountering sudden load changes, traditional methods cannot adapt in time, causing equipment overload or failure, affecting the efficiency and life of the entire organization. Summary of the invention
[0005] The purpose of the present invention is to solve the shortcomings in the prior art and to propose a load analysis method for an intelligent hoisting mechanism.
[0006] In order to achieve the above object, the present invention adopts the following technical solution: a load analysis method of an intelligent hoisting mechanism, comprising the following steps:
[0007] S1: Based on the working cycle of the winch mechanism and the characteristics of load changes, load data is collected, time windows are divided, and data in multiple time periods are clustered and analyzed to obtain a multi-scale cluster analysis framework;
[0008] S2: Based on the multi-scale cluster analysis framework, analyze the load data changes in multiple time periods, identify the short-term, medium-term and long-term load fluctuation characteristics, form the basis for load pattern recognition, and obtain a time scale load pattern set;
[0009] S3: Based on the time scale load pattern set, a cross-comparison analysis is performed on the short-term, medium-term and long-term load patterns to identify the core connections and change rules between the patterns and obtain a load pattern relationship mapping;
[0010] S4: Based on the load pattern relationship mapping, integrate the load pattern and environmental characteristics, construct a data set and classify and annotate the load pattern and characteristics, prepare standardized input data, and obtain a reinforcement learning training set;
[0011] S5: Based on the reinforcement learning training set, execute the model training process, initialize the model parameters, set the learning rate and loss function, adjust the weights and biases through the training data cycle, and obtain the deep reinforcement learning adjustment model;
[0012] S6: Based on the deep reinforcement learning adjustment model, deploy tests in the actual environment, adapt working conditions, implement load adjustment strategies, collect operating data, evaluate efficiency and energy consumption impact, and obtain an evaluation of the implementation effect of the load adjustment strategy.
[0013] As a further solution of the present invention, the multi-scale clustering analysis framework includes a time window definition, a load data clustering set, and a time scale identifier. The time scale load pattern set includes a short-term fluctuation feature set, a medium-term adjustment trend set, and a long-term change law set. The load pattern relationship mapping is specifically a pattern fitting index set, a difference description set, and a time scale conversion relationship set. The reinforcement learning training set includes an environmental state set, a load pattern identification set, and an expected response set. The deep reinforcement learning adjustment model is specifically a model structure, a parameter optimization set, and a strategy generation mechanism. The load adjustment strategy implementation effect evaluation includes adaptability test records, efficiency improvement indicators, and adjustment strategy success rate statistics.
[0014] As a further solution of the present invention, based on the multi-scale cluster analysis framework, the load data changes in multiple time periods are analyzed to identify the short-term, medium-term and long-term load fluctuation characteristics, forming the basis for load pattern recognition. The steps of obtaining the time scale load pattern set are specifically as follows:
[0015] S201: Based on the multi-scale cluster analysis framework, perform time series analysis on load data in multiple time periods, identify fluctuation patterns and periodic characteristics of loads in multiple time periods, and obtain a load fluctuation basic mapping;
[0016] S202: using the load fluctuation basic mapping, using the KMeans algorithm to perform a clustering operation, distinguishing short-term, medium-term, and long-term load change trends, extracting fluctuation characteristics of each time scale, and obtaining a time scale feature set;
[0017] S203: Perform pattern recognition through the time scale feature set, collect and analyze time series data, extract key time features and classify them, assign a time label to each load pattern, and obtain a time scale load pattern set.
[0018] As a further solution of the present invention, the KMeans algorithm adopts the formula
[0019]
[0020] Compute the clustering of the load fluctuation base map data, where is the number of clusters, representing short-term, medium-term, and long-term, For the The set of data points in a cluster, is the data point, For the The center point of the cluster, is the distance weight from the data point to the cluster center, is the intra-cluster variation weight, For the The degree of dispersion of the internal data points of a cluster, is the distance weight between clusters, For the The average distance between a cluster and other cluster centers, is the time scale impact weight, For the The time scale influencing factor of each cluster.
[0021] As a further solution of the present invention, based on the time scale load pattern set, the short-term, medium-term and long-term load patterns are cross-compared and analyzed to identify the core connections and change rules between the patterns, and the steps of obtaining the load pattern relationship mapping are specifically as follows:
[0022] S301: performing a preliminary cluster analysis on the time scale load pattern set, identifying short-term, medium-term and long-term load patterns by screening the hoisting mechanism load data in multiple time scales, marking key categories, and comparing the wire rope and winch load characteristics to obtain the load pattern cluster distribution;
[0023] S302: Based on the cluster distribution of the load pattern, tracking the tension change, identifying the growth and reduction trends, comparing the change ranges between the hook and the load limiter, and obtaining the load pattern change trend analysis results;
[0024] S303: Based on the load pattern change trend analysis result, analyze the interactive relationship between the winch and the wire rope, draw a pattern change connection diagram, mark the conversion path and influencing factors, and obtain the load pattern relationship identification result between time scales;
[0025] S304: Based on the load pattern relationship identification result between the time scales, analyze the load pattern characteristics, mark each pattern, integrate key change rules, and obtain a load pattern relationship mapping.
[0026] As a further solution of the present invention, based on the load pattern relationship mapping, integrating the load pattern and the environmental characteristics, constructing a data set and classifying and labeling the load pattern and the characteristics, preparing standardized input data, and obtaining the reinforcement learning training set are specifically as follows:
[0027] S401: Based on the load mode relationship mapping, a temperature and humidity sensor is used to collect environmental information. For the wire rope and winch components in the hoisting mechanism, the obtained information is transmitted to the central processing unit in real time through the data interface. Each piece of data is marked with a timestamp, and the tension of the wire rope and the speed of the winch are monitored in real time. The monitoring data is classified and sorted according to the time series to obtain the original data set;
[0028] S402: using the original data set, using a frequency analysis method, analyzing the interaction between the wire rope tension and the winch speed, determining the pattern combination through statistical analysis, identifying the change trend under differentiated load conditions, and obtaining a pattern recognition data set;
[0029] S403: Segment and label the data through the pattern recognition data set, extract and vectorize the features according to the operation characteristics of the hook and the load limiter, integrate the feature vector and labeling information, match the target scene data set of reinforcement learning, and obtain the reinforcement learning training set.
[0030] As a further solution of the present invention, based on the reinforcement learning training set, the model training process is executed, the model parameters are initialized, the learning rate and the loss function are set, and the weights and biases are adjusted through the training data cycle to obtain the steps of the deep reinforcement learning adjustment model:
[0031] S501: Based on the reinforcement learning training set, randomly select starting parameters, set initial learning rate, start model training process, including allocating memory to store model parameters, calling processor to perform preliminary calculation, and obtain initial parameter model;
[0032] S502: Based on the initial parameter model, a batch of training data is selected, a cross entropy loss function is used, the loss function value is calculated, and the model parameters are updated using the gradient descent method, including calculating the gradient of each parameter, adjusting the parameter value according to the gradient direction, and looping until the current batch of training data is processed, and a parameter adjustment model is generated;
[0033] S503: Based on the parameter adjustment model, continue to iterate and use the remaining training data to evaluate the model performance indicators, and stop when the improvement is lower than the threshold or reaches the maximum number of iterations to obtain the deep reinforcement learning adjustment model.
[0034] As a further solution of the present invention, the cross entropy loss function adopts the formula
[0035]
[0036] in, is the improved loss value, is the one-hot representation of the true label, is the prior probability of the label, is the adjustment factor, is the probability distribution predicted by the model, is a small constant, is the weight magnification factor.
[0037] Compared with the prior art, the advantages and positive effects of the present invention are:
[0038] In this invention, the load changes are finely divided through a multi-scale cluster analysis framework, revealing the short-term to long-term load fluctuation characteristics, providing a deeper data analysis for the hoisting mechanism. Through load pattern relationship mapping and deep reinforcement learning technology, dynamic adaptation and optimization control of load conditions are achieved. This technological advancement promotes energy efficiency optimization, reduces energy waste, and significantly improves fault prevention capabilities, enhancing the stable operation and long-term reliability of the hoisting mechanism. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 It is a schematic diagram of the main steps of the present invention;
[0040] Figure 2 It is a schematic diagram of the refinement of S1 of the present invention;
[0041] Figure 3 It is a schematic diagram of the refinement of S2 of the present invention;
[0042] Figure 4 It is a schematic diagram of the refinement of S3 of the present invention;
[0043] Figure 5 It is a schematic diagram of the refinement of S4 of the present invention;
[0044] Figure 6 It is a detailed schematic diagram of S5 of the present invention;
[0045] Figure 7 It is a detailed schematic diagram of S6 of the present invention. DETAILED DESCRIPTION
[0046] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.
[0047] In the description of the present invention, it should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, in the description of the present invention, "multiple" means two or more, unless otherwise clearly and specifically defined. Example
[0048] See also Figure 1 The present invention provides a technical solution: a load analysis method for an intelligent hoisting mechanism, comprising the following steps:
[0049] S1: Based on the working cycle of the winch mechanism and the characteristics of load changes, load data is collected, time windows are divided, and data in multiple time periods are clustered and analyzed to obtain a multi-scale cluster analysis framework;
[0050] S2: Based on the multi-scale cluster analysis framework, the load data changes in multiple time periods are analyzed to identify the short-term, medium-term and long-term load fluctuation characteristics, forming the basis for load pattern recognition and obtaining the time scale load pattern set;
[0051] S3: Based on the time scale load pattern set, a cross-comparison analysis is performed on the short-term, medium-term and long-term load patterns to identify the core connections and change rules between the patterns and obtain the load pattern relationship mapping;
[0052] S4: Based on the load pattern relationship mapping, integrate the load pattern and environmental characteristics, build a data set and classify and annotate the load pattern and characteristics, prepare standardized input data, and obtain the reinforcement learning training set;
[0053] S5: Based on the reinforcement learning training set, execute the model training process, initialize the model parameters, set the learning rate and loss function, adjust the weights and biases through the training data cycle, and obtain the deep reinforcement learning adjustment model;
[0054] S6: Based on the deep reinforcement learning adjustment model, deploy the test in the actual environment, adapt the working conditions, implement the load adjustment strategy, collect operation data, evaluate the efficiency and energy consumption impact, and obtain the evaluation of the implementation effect of the load adjustment strategy.
[0055] The multi-scale clustering analysis framework includes time window definition, load data clustering set, and time scale identifier. The time scale load pattern set includes short-term fluctuation feature set, medium-term adjustment trend set, and long-term change law set. The load pattern relationship mapping is specifically a pattern fitting indicator set, a difference description set, and a time scale conversion relationship set. The reinforcement learning training set includes an environmental state set, a load pattern identification set, and an expected response set. The deep reinforcement learning adjustment model is specifically a model structure, a parameter optimization set, and a strategy generation mechanism. The evaluation of the implementation effect of the load adjustment strategy includes adaptability test records, efficiency improvement indicators, and adjustment strategy success rate statistics.
[0056] See also Figure 2 Based on the working cycle of the winch mechanism and the characteristics of load changes, the load data is collected, the time window is divided, and the data in multiple time periods are clustered and analyzed. The specific steps of obtaining the multi-scale cluster analysis framework are as follows:
[0057] S101: Based on the working cycle and load change characteristics of the hoisting mechanism, record the load value and change trend, pay attention to the conversion between high load and low load state, and generate a load data set;
[0058] Based on the working cycle and load change characteristics of the winch mechanism, the load values and change trends are recorded. The time series analysis method is adopted, and the Pandas library in Python is used to perform time series processing on the load data. The DataFrame.resample() method is used to resample the data at time intervals in order to capture the transition between high load and low load states and generate a load data set.
[0059] S102: based on the load data set, select key time points of load change, divide the data set of the key time points into independent time periods, reflect the differentiated load change pattern, and generate a segmented load feature set;
[0060] Based on the load data set, the key time points of load change are selected, and the change point detection algorithm is adopted. The ruptures library in Python is used to perform change point detection. The ruptures.Detect() function is used to identify significant change points in the load data. The detection algorithm is set to the Pelt model, and the minimum number of change points is determined to be min_size=10. In this way, the data set is divided into independent time periods to reflect the differentiated load change pattern and generate a segmented load feature set.
[0061] S103: Based on the segmented load feature set, compare the fit of data in multiple time periods, classify time periods of similar load change patterns, construct a multi-dimensional feature vector based on the classification results, apply the feature vector to perform cluster analysis, establish representative features of multiple types of load change patterns, and generate a multi-scale cluster analysis framework.
[0062] Based on the segmented load feature set, the fit of the data in multiple time periods is compared. A multidimensional vector space model is adopted, and the feature vector is constructed using the scikit-learn library in Python. sklearn.decomposition.PCA() is used for principal component analysis to reduce the dimension, and n_components=5 is set as the target dimension. Then sklearn.cluster.KMeans is applied for cluster analysis, and the number of clusters is set to n_clusters=3 to determine the representative characteristics of multiple types of load change patterns and generate a multi-scale clustering analysis framework.
[0063] See also Figure 3 ,Based on the multi-scale cluster analysis framework, the load data changes in multiple time periods are analyzed,,the short-term, medium-term and long-term load fluctuation characteristics are identified,,forming the basis for load pattern recognition. The specific steps for obtaining the time scale load pattern set are:
[0064] S201: Based on the multi-scale cluster analysis framework, time series analysis is performed on the load data in multiple time periods to identify the fluctuation patterns and periodic characteristics of the loads in multiple time periods and obtain a load fluctuation basic mapping;
[0065] Based on the multi-scale clustering analysis framework, time series analysis is performed on load data in multiple time periods to identify the fluctuation patterns and periodic characteristics of loads in multiple time periods, and the basic mapping of load fluctuations is obtained. The time series analysis method is adopted, and the Python Pandas library is used for data processing. The timestamp is parsed through the datetime module, and the time window is set to daily. The groupby method is used to group the data in different time periods, and the rolling and mean functions are used to calculate the moving average in each group to smooth the time series data. The matplotlib library is used to draw time series charts to show a visual representation of load fluctuations, and finally a basic mapping of load fluctuations is generated.
[0066] S202: using the load fluctuation basic mapping, using the KMeans algorithm to perform clustering operations, distinguishing short-term, medium-term, and long-term load change trends, extracting the fluctuation characteristics of each time scale, and obtaining a time scale feature set;
[0067] The load fluctuation basic mapping is used to perform clustering operations, distinguish short-term, medium-term, and long-term load change trends, extract the fluctuation characteristics of each time scale, and obtain the time scale feature set. The clustering algorithm is implemented using the KMeans algorithm in the Python Scikit-learn library. The cluster number parameter n_clusters is set to 3, representing short-term, medium-term, and long-term load changes. The load fluctuation basic mapping data is clustered using the fit_predict method. According to the clustering results, the average value of the data points in each class is calculated through the numpy library to extract the fluctuation characteristics of each time scale, and finally generate the time scale feature set.
[0068] The KMeans algorithm uses the formula
[0069]
[0070] Compute the clustering of the load fluctuation base map data, where is the number of clusters, representing short-term, medium-term, and long-term, For the The set of data points in a cluster, is the data point, For the The center point of the cluster, is the distance weight from the data point to the cluster center, is the intra-cluster variation weight, For the The degree of dispersion of the internal data points of a cluster, is the distance weight between clusters, For the The average distance between a cluster and other cluster centers, is the time scale impact weight, For the The time scale influencing factor of each cluster.
[0071] First select Initial centers are assigned to each data point, and then the nearest center is assigned, taking into account the distance from the data point to the center, the intra-cluster variability, the average distance between clusters, and the influence of the time scale. The center of each cluster is then recalculated, taking into account the setting of the weight coefficient. , , , The value of is determined by cross-validation or an optimization algorithm based on historical data, and the process is repeated until the convergence condition is met.
[0072] S203: Perform pattern recognition through the time scale feature set, collect and analyze time series data, extract key time features and classify them, assign a time label to each load pattern, and obtain a time scale load pattern set.
[0073] Through the time scale feature set, pattern recognition is performed, time series data is collected and analyzed, key time features are extracted and classified, time labels are assigned to each load pattern, and the time scale load pattern set is obtained. Pattern recognition technology is used to perform classification operations using the DecisionTreeClassifier decision tree classifier in Python's Scikit-learn library. The decision tree depth max_depth parameter is set to 5 to control the model complexity. The time scale feature set is learned using the fit method, and the new time series data is classified using the predict method. Based on the classification results, different time labels are extracted from the Pandas library using the unique function to finally generate a time scale load pattern set.
[0074] See also Figure 4 ,Based on the time scale load pattern set, a cross-comparison analysis is conducted on the short-term, medium-term and long-term load patterns to identify the core connections and change rules between the patterns.,The specific steps for obtaining the load pattern relationship mapping are:
[0075] S301: Perform a preliminary cluster analysis on the time scale load pattern set, identify short-term, medium-term and long-term load patterns by screening the hoisting mechanism load data in multiple time scales, identify key categories, compare the wire rope and winch load characteristics, and obtain the load pattern cluster distribution;
[0076] A preliminary cluster analysis of the time scale load pattern set was performed using the K-means clustering algorithm. The KMeans function was executed using the scikit-learn library, and the n_clusters parameter was specified to identify the winch load data within different time scales. The initial centroid selection and the number of repeated algorithm runs were optimized by setting init='k-means++' and n_init=10 to identify short-term, medium-term, and long-term load patterns, identify key categories, and compare the load characteristics of wire ropes and winches to obtain the cluster distribution of load patterns.
[0077] S302: Based on the cluster distribution of the load pattern, track the tension change, identify the growth and reduction trends, compare the change ranges between the hook and the load limiter, and obtain the load pattern change trend analysis results;
[0078] Based on the cluster distribution of load patterns, the tension changes are tracked. The time series analysis method is adopted, the Pandas library is used for data processing, the rolling() function is executed, the window parameter is specified to smooth the data and track the trend changes, the growth and decrease trends are identified, the change amplitudes between the hook and the load limiter are compared, and the load pattern change trend analysis results are obtained.
[0079] S303: Based on the load pattern change trend analysis results, the interactive relationship between the winch and the wire rope is analyzed, a pattern change connection diagram is drawn, the conversion path and the influencing factors are marked, and the load pattern relationship identification results between time scales are obtained;
[0080] Based on the analysis results of load pattern change trend, the interactive relationship between the winch and the wire rope is analyzed. The graph theory analysis method is adopted, and the network diagram is constructed using the NetworkX library. The add_node() and add_edge() functions are executed to add nodes and edges, respectively, to simulate the connection between the winch and the wire rope. The mode change connection diagram is drawn, the conversion path and influencing factors are marked, and the load pattern relationship identification results between time scales are obtained.
[0081] S304: Based on the load pattern relationship identification results between time scales, analyze the load pattern characteristics, mark each pattern, integrate key change rules, and obtain the load pattern relationship mapping.
[0082] Based on the load pattern relationship identification results between time scales, the load pattern characteristics are analyzed, and the principal component analysis (PCA) method is adopted. The PCA function is executed using the scikit-learn library, and the n_components parameter is specified to extract the main features and integrate the key change rules to obtain the load pattern relationship mapping.
[0083] See also Figure 5 ,Based on the load pattern relationship mapping, the load pattern and environmental characteristics are integrated, the data set is constructed and the load pattern and characteristics are classified and labeled, and the standardized input data is prepared. The specific steps to obtain the reinforcement learning training set are:
[0084] S401: Based on the load mode relationship mapping, the temperature and humidity sensors are used to collect environmental information. For the wire rope and winch components in the hoisting mechanism, the obtained information is transmitted to the central processing unit in real time through the data interface. Each piece of data is marked with a timestamp to monitor the tension of the wire rope and the speed of the winch in real time. The monitoring data is classified and sorted according to the time series to obtain the original data set;
[0085] Based on load mode relationship mapping, temperature and humidity sensors are used to collect environmental information. For the wire rope and winch components in the hoisting mechanism, the obtained information is transmitted to the central processing unit in real time through the data interface. Each data is marked with a timestamp, and the tension of the wire rope and the speed of the winch are monitored in real time. The Python Pandas library is used to time serialize the received data, and each data with a timestamp is stored in the DataFrame structure. The sort_values method is used to sort according to the timestamp to ensure that the data is sorted in chronological order to obtain the original data set.
[0086] S402: using the original data set, using the frequency analysis method, analyzing the interaction between the wire rope tension and the winch speed, determining the pattern combination through statistical analysis, identifying the change trend under differentiated load conditions, and obtaining the pattern recognition data set;
[0087] The original data set is used to analyze the interaction between wire rope tension and winch speed by frequency analysis method. The NumPy library is used for numerical analysis of the data. The time series data is fast Fourier transformed by the fft function to analyze the frequency characteristics of wire rope tension and winch speed. The logical_and function is used in combination with conditional screening to identify the pattern combination at a specific frequency and identify the change trend under differentiated load conditions. Through this process, a pattern recognition data set is obtained.
[0088] S403: Segment and label the data through the pattern recognition data set, extract and vectorize the features according to the operation characteristics of the hook and the load limiter, integrate the feature vector and labeling information, match the target scene data set of the reinforcement learning, and obtain the reinforcement learning training set.
[0089] Through the pattern recognition data set, the data is segmented and labeled, and feature extraction and vectorization are performed on the operating characteristics of the hook and load limiter. The features are vectorized using the TfidfVectorizer function in the scikit-learn library, and different operating characteristics are labeled with LabelEncoder to generate digital labels. The feature vectors and labeling information are integrated, and the merge function of pandas is used to merge the feature vector DataFrame with the labeling information DataFrame to provide training data for the reinforcement learning model, match the target scenario data set of the reinforcement learning, and obtain the reinforcement learning training set.
[0090] See also Figure 6 , based on the reinforcement learning training set, execute the model training process, initialize the model parameters, set the learning rate and loss function, adjust the weights and biases through the training data cycle, and obtain the steps of the deep reinforcement learning adjustment model as follows:
[0091] S501: Based on the reinforcement learning training set, randomly select starting parameters, set the initial learning rate, and start the model training process, including allocating memory to store model parameters, calling the processor to perform preliminary calculations, and obtaining an initial parameter model;
[0092] Based on the reinforcement learning training set, the starting parameters are randomly selected, and the uniform distribution method is used to generate an initial value for each model parameter in the range [-0.05, 0.05]. The initial learning rate is set to 0.01, and the TensorFlow framework is used to allocate memory to store the model parameters. The CPU or GPU processor is called for preliminary calculations, and the initial allocation of parameters is activated by executing the initialize_all_variables() method of TensorFlow, and the initial parameter model is obtained by simple forward propagation calculation.
[0093] S502: Based on the initial parameter model, a batch of training data is selected, a cross entropy loss function is used, the loss function value is calculated, and the model parameters are updated using the gradient descent method, including calculating the gradient of each parameter, adjusting the parameter value according to the gradient direction, and looping until the current batch of training data is processed and a parameter adjustment model is generated;
[0094] Based on the initial parameter model, a batch of training data is selected, the loss function value is calculated, and the cross entropy loss function is used to calculate the difference between the output and the true label through TensorFlow's tf.nn.softmax_cross_entropy_with_logits() method. The model parameters are updated using the gradient descent method, including calculating the gradient of each parameter, using TensorFlow's tf.gradients() function to obtain gradient information, adjusting the parameter value according to the gradient direction, and using the Adam optimizer with a learning rate of 0.01. The gradient update is applied through TensorFlow's tf.train.AdamOptimizer().apply_gradients() method. The cycle continues until the current batch of training data is processed and a parameter adjustment model is generated.
[0095] The cross entropy loss function uses the formula
[0096]
[0097] in, is the improved loss value, is the one-hot representation of the true label, is the prior probability of the label, is the adjustment factor, is the probability distribution predicted by the model, is a small constant, is the weight magnification factor.
[0098] First, the one-hot representation of the true label The corresponding prior probability By introducing the adjustment factor , which can give different importance to labels of different categories, which helps to improve the accuracy of the model in unbalanced datasets. Then, the probability distribution of the model prediction Introducing a small constant , to avoid division by zero when calculating logarithms. In addition, by introducing the weight magnification factor , can strengthen or weaken the impact of prediction probability, thereby controlling the sensitivity of the model to uncertainty to a certain extent. Ultimately, these improvements enable the loss function to more finely adjust the learning path of the model during training, optimize the model's adaptability and generalization performance to complex data features, and confirm the coefficient value and The specific method steps include performing multiple experiments using a validation set and selecting those coefficient values that optimize the performance on the validation set. For example, this can be achieved through grid search or Bayesian optimization methods.
[0099] S503: Based on the parameter adjustment model, continue to iterate and use the remaining training data to evaluate the model performance indicators. When the improvement is lower than the threshold or reaches the maximum number of iterations, stop and obtain the deep reinforcement learning adjustment model.
[0100] Based on the parameter adjustment model, continue to iterate using the remaining training data, evaluate the model performance indicators, and stop when the improvement is lower than the threshold or reaches the maximum number of iterations. Use accuracy as the performance indicator and evaluate by calculating the ratio of the number of correct predictions of the model on the validation set to the total number of predictions. Set the threshold to 1% improvement and the maximum number of iterations to 10,000. Use TensorFlow's tf.metrics.accuracy method to calculate the accuracy. Determine whether to continue iterating based on the improvement in accuracy and the number of iterations to obtain a deep reinforcement learning adjustment model.
[0101] See also Figure 7 ,Based on the deep reinforcement learning adjustment model, the test is deployed in the actual environment, the working conditions are adapted, the load adjustment strategy is implemented, the operation data is collected, the efficiency and energy consumption impact are evaluated, and the specific steps for evaluating the implementation effect of the load adjustment strategy are as follows:
[0102] S601: Adjust the model based on deep reinforcement learning, deploy the test in the actual environment, use the sensor array to record the model response, use the environmental simulator to set differentiated working conditions, monitor and record the performance of the model under multiple conditions, and generate environmental adaptation data;
[0103] Based on deep reinforcement learning, the model is adjusted and deployed for testing in the actual environment. The sensor array is used to record the model response, and the environmental simulator is used to set differentiated working conditions. LabVIEW software is used to communicate with the hardware interface. Different temperature, humidity and load conditions are set through the program to simulate the actual working environment. The performance of the model under multiple conditions is monitored and recorded. Matlab is used for data analysis to sort out the model response time and performance indicators and generate environmental adaptation data.
[0104] S602: Using the environment adaptation data, identify the bottlenecks of response time delay and processing capacity through Apache Spark, adjust the deep network layers and the number of neurons, use the simulated load generator to simulate the actual operating conditions, record the operating data before and after the model adjustment, and obtain the strategy optimization data;
[0105] Using the environment adaptation data, Apache Spark is used to identify response time delays and processing capacity bottlenecks. Spark's big data processing capabilities are used to perform batch analysis on the collected data, identify the delays and bottlenecks of the model when processing different types of data, adjust the deep network layers and the number of neurons, and use a simulated load generator written in Python to simulate actual operating conditions. The deep network structure and parameters of the model are adjusted through TensorFlow, and the operating data before and after the model adjustment is recorded. Jupyter Notebook is used for data visualization and comparison to obtain strategy optimization data.
[0106] S603: Analyze the strategy optimization data, use performance evaluation tools to compare the processing speed, accuracy and system energy consumption before and after optimization, evaluate the effectiveness of model optimization, and obtain the evaluation of the implementation effect of the load adjustment strategy.
[0107] Analyze the strategy optimization data and use performance evaluation tools, specifically the classification_report and energy_usage packages in the Python library Scikit-learn, to evaluate the model's processing speed, accuracy, and system energy consumption. Use these tools to analyze the performance differences of the models before and after optimization, compare the model performance under different parameter settings, and the changes in energy consumption under the same load conditions, evaluate the effectiveness of model optimization, use the obtained data to generate an evaluation report, and obtain an evaluation of the implementation effect of the load adjustment strategy.
[0108] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.
Claims
1. A method for analyzing load of an intelligent hoisting mechanism, characterized in that: The following steps are involved: Based on the working cycle of the winch mechanism and the characteristics of load changes, load data is collected, time windows are divided, and data in multiple time periods are clustered and analyzed to obtain a multi-scale cluster analysis framework; The multi-scale cluster analysis framework includes time window definition, load data clustering set, and time scale identifier; Based on the multi-scale cluster analysis framework, the load data changes in multiple time periods are analyzed to identify the short-term, medium-term and long-term load fluctuation characteristics, forming the basis for load pattern recognition and obtaining a time scale load pattern set; The time scale load pattern set includes a short-term fluctuation feature set, a mid-term adjustment trend set, and a long-term change law set; The steps of the time scale load pattern set are specifically: Based on the multi-scale cluster analysis framework, time series analysis is performed on load data in multiple time periods to identify fluctuation patterns and periodic characteristics of loads in multiple time periods and obtain a load fluctuation base map; The load fluctuation basic mapping is adopted, and the KMeans algorithm is used to perform clustering operations, distinguish short-term, medium-term, and long-term load change trends, extract the fluctuation characteristics of each time scale, and obtain a time scale feature set; By using the time scale feature set, pattern recognition is performed, time series data is collected and analyzed, key time features are extracted and classified, a time label is assigned to each load pattern, and a time scale load pattern set is obtained; Based on the time scale load pattern set, a cross-comparison analysis is performed on the short-term, medium-term and long-term load patterns to identify the core connections and change rules between the patterns and obtain a load pattern relationship mapping; The load pattern relationship mapping specifically includes a pattern fit index set, a difference description set, and a time scale conversion relationship set; The steps of load mode relationship mapping are specifically as follows: Performing a preliminary cluster analysis on the time scale load pattern set, identifying short-term, medium-term, and long-term load patterns by screening the winch load data in multiple time scales, identifying key categories, and comparing the wire rope and winch load characteristics to obtain the load pattern cluster distribution; Based on the cluster distribution of the load pattern, track the tension change, identify the growth and reduction trends, compare the change amplitudes between the hook and the load limiter, and obtain the load pattern change trend analysis results; Based on the load pattern change trend analysis results, the interactive relationship between the winch and the wire rope is analyzed, a pattern change connection diagram is drawn, the conversion path and influencing factors are marked, and the load pattern relationship identification results between time scales are obtained; Based on the load pattern relationship identification results between the time scales, the load pattern characteristics are analyzed, each pattern is marked, and key change rules are integrated to obtain a load pattern relationship mapping; Based on the load pattern relationship mapping, the load pattern and the environmental characteristics are integrated, a data set is constructed, and the load pattern and the characteristics are classified and labeled, and standardized input data is prepared to obtain a reinforcement learning training set; The reinforcement learning training set includes an environment state set, a load mode identification set, and an expected response set; The steps of the reinforcement learning training set are specifically as follows: Based on the load mode relationship mapping, a temperature and humidity sensor is used to collect environmental information. For the wire rope and winch components in the hoisting mechanism, the obtained information is transmitted to the central processing unit in real time through the data interface. Each piece of data is marked with a timestamp to monitor the tension of the wire rope and the speed of the winch in real time. The monitoring data is classified and sorted according to the time series to obtain the original data set; Using the original data set, the interaction between the wire rope tension and the winch speed is analyzed by frequency analysis method, the pattern combination is determined by statistical analysis, the change trend under differentiated load conditions is identified, and the pattern recognition data set is obtained; The data is segmented and labeled through the pattern recognition data set, feature extraction and vectorization are performed for the operating characteristics of the hook and the load limiter, the feature vector and the labeling information are integrated, and the target scene data set of the reinforcement learning is matched to obtain the reinforcement learning training set; Based on the reinforcement learning training set, a model training process is performed, model parameters are initialized, a learning rate and a loss function are set, and weights and biases are adjusted through a training data cycle to obtain a deep reinforcement learning adjustment model; The deep reinforcement learning adjustment model specifically includes a model structure, a parameter optimization set, and a strategy generation mechanism; Based on the deep reinforcement learning adjustment model, the test is deployed in the actual environment, the working conditions are adapted, the load adjustment strategy is implemented, the operation data is collected, the efficiency and energy consumption impact are evaluated, and the evaluation of the implementation effect of the load adjustment strategy is obtained; The load adjustment strategy implementation effect evaluation includes adaptability test records, efficiency improvement indicators, and adjustment strategy success rate statistics.
2. The intelligent hoisting mechanism load analysis method according to claim 1, characterized in that: The KMeans algorithm uses the formula ; Compute the clustering of the load fluctuation base map data, where is the number of clusters, representing short-term, medium-term, and long-term, For the The set of data points in a cluster, is the data point, For the The center point of the cluster, is the distance weight from the data point to the cluster center, is the intra-cluster variation weight, For the The degree of dispersion of the internal data points of a cluster, is the distance weight between clusters, For the The average distance between a cluster and other cluster centers, is the time scale impact weight, For the The time scale influencing factor of each cluster.
3. The intelligent hoisting mechanism load analysis method according to claim 1, characterized in that: Based on the reinforcement learning training set, the model training process is executed, the model parameters are initialized, the learning rate and the loss function are set, and the weights and biases are adjusted through the training data cycle to obtain the steps of the deep reinforcement learning adjustment model: Based on the reinforcement learning training set, randomly selecting starting parameters, setting an initial learning rate, and starting a model training process, including allocating memory to store model parameters, calling a processor to perform preliminary calculations, and obtaining an initial parameter model; Based on the initial parameter model, a batch of training data is selected, a cross entropy loss function is used, the loss function value is calculated, and the model parameters are updated using the gradient descent method, including calculating the gradient of each parameter, adjusting the parameter value according to the gradient direction, and looping until the current batch of training data is processed, and a parameter adjustment model is generated; Based on the parameter adjustment model, continue to iterate and use the remaining training data to evaluate the model performance indicators, and stop when the improvement is lower than the threshold or reaches the maximum number of iterations to obtain the deep reinforcement learning adjustment model.
4. The intelligent hoisting mechanism load analysis method according to claim 3 is characterized in that: The cross entropy loss function uses the formula ; in, is the improved loss value, is the one-hot representation of the true label, is the prior probability of the label, is the adjustment factor, is the probability distribution predicted by the model, is a small constant, is the weight magnification factor.
Citation Information
Patent Citations
Reactive load prediction method based on K-means clustering and random forest algorithm
CN116307111A
Intelligent power module management method and system
CN117454093A
Green intelligent computing center energy efficiency optimization method and system
CN117539726A
Building engineering quality real-time monitoring system
CN117611015A
Cooperative operation control method and device of power grid equipment, equipment and storage medium
CN117895579A
