Iron tower safety early warning method and system based on model analysis

Abstract

The embodiment of the invention provides an iron tower safety early warning method and system based on model analysis, and relates to the technical field of artificial intelligence. A composite training mechanism (a first model attribute tag) of a structure cooperative training mode and a data cooperative training mode or a training mechanism (a second model attribute tag) of only the data cooperative training mode is adopted, and the model training process is optimized. By segmenting a model architecture to a plurality of processing units (structure collaboration) and distributing sample data to a plurality of processing units (data collaboration), efficient parallel processing is realized, and network parameter learning is accelerated. And finally, by using the generated target iron tower safety early warning network model, any input iron tower associated characteristic data can be accurately analyzed, a safety early warning result is generated in time, and the efficiency and reliability of iron tower safety management are remarkably improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and more specifically, to a tower safety early warning method and system based on model analysis. Background Art

[0002] In the fields of electricity, communications, etc., iron towers are important infrastructures, and their safety performance is directly related to the stability of the network and the reliability of services. However, since iron towers are in outdoor environments for a long time, they are susceptible to wind and rain erosion, temperature changes and other factors, which poses a safety hazard. Therefore, it is particularly important to conduct real-time monitoring and early warning of the safety status of iron towers.

[0003] Traditional tower safety monitoring methods mainly rely on manual inspections and simple sensor monitoring, which have many shortcomings. Manual inspections are time-consuming and laborious, and are limited by personnel experience and skill levels, making it difficult to discover all potential safety issues. Although simple sensor monitoring can obtain some basic data in real time, its data processing and analysis capabilities are limited, making it difficult to accurately judge the safety status of the tower. Summary of the invention

[0004] In order to at least overcome the above-mentioned deficiencies in the prior art, the purpose of the present application is to provide a tower safety early warning method and system based on model analysis.

[0005] In a first aspect, the present application provides a tower safety early warning method based on model analysis, which is applied to a tower safety early warning system based on model analysis. The method comprises: Obtaining model knowledge features of a tower safety warning network model for which network parameter learning is to be performed; Based on the model knowledge characteristics, determining the model attribute label of the iron tower safety early warning network model; If the iron tower safety warning network model matches the first model attribute label, network parameter learning is performed on the iron tower safety warning network model according to a first collaborative training mechanism, wherein the first collaborative training mechanism is used to represent a composite training mechanism of a structure collaborative training mode and a data collaborative training mode; If the tower safety warning network model matches the second model attribute label, network parameter learning is performed on the tower safety warning network model according to the second collaborative training mechanism to generate a target tower safety warning network model, and the second collaborative training mechanism is used to represent a training mechanism of a data collaborative training mode; wherein the structure collaborative training mode is used to represent a training mode in which the model architecture of the tower safety warning network model is divided into multiple model processing units for network parameter learning, and the data collaborative training mode is used to represent a training mode in which the sample tower-related feature data of the tower safety warning network model is distributed to multiple model processing units for network parameter learning; The target tower safety warning network model is used to analyze the tower-related characteristic data of any input target tower to generate a corresponding safety warning result.

[0006] In the second aspect, an embodiment of the present application also provides a tower safety warning system based on model analysis, the tower safety warning system based on model analysis includes a processor and a machine-readable storage medium, the machine-readable storage medium stores a computer program, and the computer program is loaded and executed in conjunction with the processor to implement the tower safety warning method based on model analysis of the first aspect above.

[0007] In combination with the above aspects, the embodiment of the present application obtains the model knowledge characteristics of the tower safety warning network model, and selects the corresponding collaborative training mechanism for network parameter learning based on the model attribute label, thereby effectively improving the accuracy and adaptability of the warning model. Specifically, for different model attribute labels, a composite training mechanism of the structure collaborative training mode and the data collaborative training mode (first model attribute label) or a training mechanism of only the data collaborative training mode (second model attribute label) is adopted to optimize the model training process. By dividing the model architecture into multiple processing units (structural collaboration) and allocating sample data to multiple processing units (data collaboration), efficient parallel processing is achieved and network parameter learning is accelerated. Finally, using the generated target tower safety warning network model, it is possible to accurately analyze any input tower-related feature data, generate safety warning results in a timely manner, and significantly improve the efficiency and reliability of tower safety management. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required to be used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0009] Figure 1 A schematic diagram of a process flow of a tower safety early warning method based on model analysis provided in an embodiment of the present application; Figure 2 A structural schematic block diagram of a tower safety warning system based on model analysis for implementing the above-mentioned tower safety warning method based on model analysis provided in an embodiment of the present application. DETAILED DESCRIPTION

[0010] The following description is intended to enable one of ordinary skill in the art to implement and utilize the present application, and the description is provided in the context of a specific application scenario and its requirements. It will be apparent to one of ordinary skill in the art that various changes may be made to the disclosed embodiments, and that the general principles defined in the present application may be applied to other embodiments and application scenarios without departing from the principles and scope of the present application. Therefore, the present application is not limited to the described embodiments, but should be given the broadest scope equal to the claims.

[0011] Step S110, obtaining model knowledge features of the tower safety warning network model for which network parameter learning is to be performed.

[0012] In detail, the model knowledge features refer to various key information and attributes used to describe and define the tower safety warning network model. These model knowledge features not only include the structural information of the model (such as the number of layers, number of neurons, connection mode, etc.), but also cover the data features used in the model training process (such as data distribution, scale, dimension, etc.), as well as the performance indicators of the model in the previous training and verification stages (such as accuracy, recall rate, F1 value, etc.). Model knowledge features are the basis for understanding and optimizing model behavior, and are crucial for determining the training strategy and optimization direction of the model.

[0013] That is, in this embodiment, in a large communication network operation scenario, the server is responsible for managing and maintaining a large number of tower facilities. In order to ensure the safe operation of the tower, a tower safety early warning network model is constructed.

[0014] The server first obtains the model knowledge features of the tower safety warning network model. These model knowledge features may include the architecture data of the tower safety warning network model, such as the number of network layers, the number of neurons in each layer, the connection method, etc.; they may also include the training data features of the tower safety warning network model, such as the distribution of data, the scale of data, the characteristic dimension of data, etc.; they may also include the performance indicators of the tower safety warning network model, such as accuracy, recall rate, F1 value, etc. in the previous training and verification stages.

[0015] Assume that this tower safety warning network model is a convolutional neural network (CNN) architecture based on deep learning. The model architecture data obtained by the server shows that the tower safety warning network model has 10 convolutional layers, and the number of filters in each convolutional layer starts from 64 and increases in multiples of 2 until 512. There are 3 fully connected layers, with the number of neurons being 1024, 512, and 256 respectively. In addition, the server also obtains features about the training data, such as the data set contains monitoring data of 100,000 towers from different regions and under different climatic conditions. The data of each tower has 20 feature dimensions, including the height, tilt angle, material strength, wind speed, temperature, etc. of the tower. At the same time, the server also obtained the accuracy of the model on the previous small-scale test set of 85%, the recall rate of 78%, and the F1 value of 81%.

[0016] Step S120, determining the model attribute label of the tower safety early warning network model based on the model knowledge characteristics.

[0017] In detail, the model attribute label is a means of classifying and identifying models according to model knowledge characteristics. It divides models into different categories based on certain key attributes of the model (such as scale, operating environment, etc.) and assigns a unique label to each category. These labels are used to guide subsequent training and optimization processes to ensure the selection of the most appropriate training mechanism and parameter adjustment strategy. For example, according to the scale or operating environment of the model, the model can be divided into "large model" or "small model", "high-performance environment model" or "low-resource environment model", etc.

[0018] For example, in this step, the following methods may be included: Method 1: Calculate the scale of the tower safety warning network model based on the model architecture data. Assume that the scale of the tower safety warning network model is mainly measured by the number of parameters of the model. The server counts the number of parameters of the above CNN model, including the filter weights and bias terms of the convolutional layer, and the weights and bias terms of the fully connected layer. After calculation, the total number of parameters reaches 50 million. If the third threshold value set in advance is 30 million parameters, since the number of parameters of the current tower safety warning network model of 50 million is greater than the third threshold value of 30 million, the server marks the model attribute label of the tower safety warning network model as the first model attribute label.

[0019] Method 2: Based on the operating environment characteristics of the tower safety warning network model. Assume that the operating environment characteristics are divided into high computing resource environment and low computing resource environment. If the operating environment where the current server is located has a large number of high-performance computing nodes, high-speed network connections, and sufficient memory resources, it is defined as the first operating environment label. If the server is in an environment with limited computing resources, low network bandwidth, and tight memory, it is defined as the second operating environment label. In the current scenario, the environment where the server is located has a powerful computing cluster that can handle multiple large-scale tasks at the same time, which belongs to the first operating environment label. Therefore, the server marks the model attribute label of the tower safety warning network model as the first model attribute label.

[0020] Step S130, if the tower safety warning network model matches the first model attribute label, network parameter learning is performed on the tower safety warning network model according to the first collaborative training mechanism, and the first collaborative training mechanism is used to represent a composite training mechanism of a structure collaborative training mode and a data collaborative training mode.

[0021] In detail, the first collaborative training mechanism is a composite training mechanism that combines the structural collaborative training mode and the data collaborative training mode. The structural collaborative training mode divides the model architecture into multiple model blocks, and assigns these blocks to different model processing units for parallel training. The data collaborative training mode divides the sample data into multiple data sequences, and each sequence is assigned to a different model processing unit for training. The first collaborative training mechanism achieves more efficient and comprehensive model training by simultaneously utilizing these two collaborative training modes, thereby improving the accuracy and generalization ability of the model.

[0022] That is, in this embodiment, since the above-mentioned tower safety warning network model is marked as the first model attribute label, the server adopts the first collaborative training mechanism to perform network parameter learning.

[0023] The first collaborative training mechanism includes a structure collaborative training mode and a data collaborative training mode.

[0024] For the structure collaborative training mode, the server divides the above CNN model architecture into multiple model blocks. Assume that it is divided into 5 model blocks.

[0025] The server first processes the model's weight parameter information. For example, for the filter weights of the convolutional layer, the server evaluates the size and importance of these weights and generates the corresponding weight coefficients. Then, the filter weights are grouped and calculated based on the weight coefficients and 5, and the filter weights are allocated to 5 different model blocks. Each model block contains a portion of specific weight parameter information.

[0026] Next, the server collaboratively updates the weight parameter information of the five model blocks to the five model processing units. The first model processing unit loads the weight parameter information of the first model block, and so on.

[0027] For the first model processing unit, it uses the loaded model block to perform feature prediction on the first knowledge vector data corresponding to each data in the input 100,000 sample tower-related feature data, and generates the first risk knowledge vector corresponding to each data.

[0028] Then, the server fuses the five risk knowledge vectors corresponding to each of the 100,000 sample tower-related feature data, for example, by weighted averaging or taking the maximum value, to generate a first risk warning signal corresponding to each data.

[0029] For the data collaborative training mode, the server divides the above 100,000 sample tower-related feature data into 5 sample feature data sequences.

[0030] The first model processing unit is responsible for processing the first sample feature data sequence. It predicts the second knowledge vector data corresponding to each sample tower-related feature data in the sequence according to the second model architecture, and generates a second risk warning signal corresponding to each data in the sequence.

[0031] The server fuses the second risk warning signals corresponding to each sample tower-associated feature data in the five sample feature data sequences, and generates second risk warning signals corresponding to each of the 100,000 sample tower-associated feature data.

[0032] Step S140, if the tower safety warning network model matches the second model attribute label, network parameter learning is performed on the tower safety warning network model according to the second collaborative training mechanism to generate a target tower safety warning network model, and the second collaborative training mechanism is used to represent the training mechanism of the data collaborative training mode. Among them, the structure collaborative training mode is used to represent the training mode of dividing the model architecture of the tower safety warning network model into multiple model processing units for network parameter learning, and the data collaborative training mode is used to represent the training mode of distributing the sample tower-related feature data of the tower safety warning network model to multiple model processing units for network parameter learning.

[0033] In detail, the second collaborative training mechanism only includes the data collaborative training mode and does not involve the structural collaborative training mode. It improves the parallelism and efficiency of the training process by dividing the sample data into multiple data sequences and assigning these sequences to different model processing units for training. Compared with the first collaborative training mechanism, the second collaborative training mechanism is more suitable for models with smaller scale or limited computing resources, and can achieve rapid optimization of the model under limited resources.

[0034] The structural collaborative training mode is a training mode that divides a complex model architecture into multiple model blocks and assigns these blocks to different model processing units for parallel training. By assigning different parts of the model to different processing units, the structural collaborative training mode can fully utilize the computing power of multi-core processors and accelerate the model training process. At the same time, since each processing unit is only responsible for a part of the model, this training mode also helps to improve the stability and reliability of training.

[0035] The data collaborative training mode is a training mode that divides sample data into multiple data sequences and distributes these sequences to different model processing units for training. By processing multiple data sequences in parallel, the data collaborative training mode can significantly increase the training speed and reduce the overall training time. In addition, since each processing unit processes a portion of the data independently, this training mode also helps to improve the generalization ability of the model, enabling it to better adapt to data of different types and distributions.

[0036] The target tower safety early warning network model refers to a neural network model that is trained and optimized to provide real-time early warning of the tower safety status. The model is trained based on a large amount of tower-related feature data, and can accurately identify potential safety hazards of the tower and generate corresponding safety early warning results. By using the target tower safety early warning network model, real-time monitoring and effective management of tower safety can be achieved, reducing the probability of tower failures and safety accidents.

[0037] The tower-related characteristic data refers to various monitoring data and attribute information related to the safety status of the tower. These data may include the tower's height, tilt angle, material strength, natural conditions such as wind speed, temperature, humidity, and historical maintenance records and usage status of the tower. By analyzing and processing these related characteristic data, the safety status of the tower can be fully understood, providing strong support for safety warnings.

[0038] In this embodiment, it is assumed that another iron tower safety warning network model is smaller in scale and is marked as a second model attribute label.

[0039] The server adopts the second collaborative training mechanism, first collaboratively updates the weight parameter information of the tower safety warning network model to the three model processing units, and divides the 50,000 sample tower-related feature data into three sample feature data sequences.

[0040] The first model processing unit uses the tower safety early warning network model to perform feature prediction on each sample tower-related feature data in the first sample feature data sequence, and generates a risk prediction signal corresponding to each data. Based on these risk prediction signals, the weight optimization direction value obtained by the first model processing unit is generated.

[0041] The server merges the weight optimization direction values ​​determined by the three model processing units to generate the weight optimization direction value of the entire model.

[0042] Then, based on this weight optimization direction value, the weight parameter information of the tower safety warning network model is updated, and this process is continuously iterated until the model converges or reaches a predetermined number of training rounds, and finally the target tower safety warning network model is generated.

[0043] Alternatively, in another case, the server collaboratively updates the weight parameter information of the tower safety warning network model to four model processing units, and divides 80,000 sample tower-related feature data into eight sample feature data sequences, with each model processing unit processing two sequences.

[0044] For the first model processing unit, it predicts each sample tower-related feature data in the two sample feature data sequences it is responsible for, generates a corresponding risk prediction signal, and obtains two sub-weight optimization direction values.

[0045] If the number of sample feature data sequences processed by the first model processing unit is less than a preset second threshold value (assuming it is 3), the two sub-weight optimization direction values ​​are aggregated to generate the weight optimization direction value obtained by the first model processing unit.

[0046] If the number of sample feature data sequences processed by all model processing units is not less than the second threshold value, the weight optimization direction values ​​respectively determined by the four model processing units are fused to generate the weight optimization direction value of the model.

[0047] Finally, based on this weight optimization direction value, the weight parameter information of the tower safety early warning network model is updated to generate the target tower safety early warning network model.

[0048] Step S150, using the target tower safety warning network model to analyze the tower-related characteristic data of any input target tower to generate a corresponding safety warning result.

[0049] In actual operation, new towers generate relevant monitoring data, and the server obtains the tower-related feature data of these target towers.

[0050] Assume that the tower-related feature data of the target tower include the current height, tilt angle, wind speed, temperature, and material fatigue. The server inputs the tower-related feature data of these target towers into the trained target tower safety warning network model. The target tower safety warning network model analyzes and processes these tower-related feature data, and finally outputs a safety warning result through multi-layer convolution and full connection calculation. If the safety warning result indicates that the tower is in a high-risk state, the server will immediately send an alarm to the maintenance personnel to notify them to take emergency measures for inspection and repair. If the safety warning result shows low risk or medium risk, the server will record the relevant data and continue to pay attention to the status changes of the tower according to the predetermined monitoring cycle.

[0051] In this way, the tower safety early warning network model can be effectively trained and applied to ensure the safe and stable operation of the tower.

[0052] Based on the above steps, the embodiment of the present application obtains the model knowledge characteristics of the tower safety warning network model, and selects the corresponding collaborative training mechanism for network parameter learning according to the model attribute label, thereby effectively improving the accuracy and adaptability of the warning model. Specifically, for different model attribute labels, a composite training mechanism of the structure collaborative training mode and the data collaborative training mode (first model attribute label) or a training mechanism of only the data collaborative training mode (second model attribute label) is adopted to optimize the model training process. By dividing the model architecture into multiple processing units (structural collaboration) and allocating sample data to multiple processing units (data collaboration), efficient parallel processing is achieved and network parameter learning is accelerated. Finally, using the generated target tower safety warning network model, it is possible to accurately analyze any input tower-related feature data, generate safety warning results in a timely manner, and significantly improve the efficiency and reliability of tower safety management.

[0053] In a possible implementation, the tower safety warning network model includes: a first model architecture based on the structure collaborative training mode, and a second model architecture based on the data collaborative training mode. The model learning data of the first model architecture includes the first knowledge vector data corresponding to each of the X sample tower-related feature data, the model prediction data of the first model architecture includes the first risk warning signal corresponding to each of the X sample tower-related feature data, the model learning data of the second model architecture includes the second knowledge vector data corresponding to each of the X sample tower-related feature data, and the model prediction data of the second model architecture includes the second risk warning signal corresponding to each of the X sample tower-related feature data, where X is a positive integer.

[0054] Step S130 includes: Step S131, for the first model architecture, divide the first model architecture into Y model blocks, and collaboratively update the weight parameter information of the Y model blocks to Y model processing units, the a-th model processing unit among the Y model processing units loads the weight parameter information of the a-th model block among the Y model blocks, Y is a positive integer, and a is not greater than Y.

[0055] Step S132: for the ath model processing unit, use the ath model block to perform feature prediction on the first knowledge vector data corresponding to each of the X sample tower-related feature data, and generate the ath risk knowledge vector corresponding to each of the X sample tower-related feature data, where X is a positive integer.

[0056] Step S133: fusing the Y risk knowledge vectors corresponding to the X sample tower-related feature data to generate first risk warning signals corresponding to the X sample tower-related feature data.

[0057] Step S134: for the second model architecture, the weight parameter information of the second model architecture is collaboratively updated to Z model processing units, and the X sample tower-associated feature data are divided into Z sample feature data sequences, each sample feature data sequence includes at least one sample tower-associated feature data, and the b-th model processing unit among the Z model processing units is used to process the b-th sample feature data sequence among the Z sample feature data sequences, where Z is a positive integer and b is a positive integer not greater than Z.

[0058] Step S135: For the b-th model processing unit, use the b-th model processing unit to predict the second knowledge vector data corresponding to each sample tower-associated feature data in the b-th sample feature data sequence according to the second model architecture, and generate a second risk warning signal corresponding to each sample tower-associated feature data in the b-th sample feature data sequence.

[0059] Step S136, fusing the second risk warning signals corresponding to each sample tower-associated feature data in the Z sample feature data sequences to generate second risk warning signals corresponding to each of the X sample tower-associated feature data.

[0060] In this embodiment, in this scenario, the value of X is 10000, that is, there are 10000 sample tower-related feature data.

[0061] For the first model architecture: First, the server divides the first model architecture into Y model blocks. Assume that Y is 5. The server will determine how to divide the model into 5 blocks based on the structure and characteristics of the model, as well as the analysis and evaluation of the model. For example, the server may divide the model based on the distribution of convolutional layers, the connection relationship of neurons, and other factors.

[0062] Then, the weight parameter information of the five model blocks is collaboratively updated to the five model processing units. The first model processing unit loads the weight parameter information of the first model block, and so on.

[0063] For the first model processing unit, it uses the loaded first model block to perform feature prediction on the first knowledge vector data corresponding to each of the 10,000 sample tower-related feature data. For example, for the first knowledge vector data of the first sample tower, the first model processing unit will generate the corresponding first risk knowledge vector through its internal calculation and processing logic. Similarly, for the second, third, and up to the 10,000th sample tower, such processing will be performed to generate the first risk knowledge vector corresponding to each of the 10,000 sample towers.

[0064] Then, the server will fuse the five risk knowledge vectors (generated by five model processing units) corresponding to each of the 10,000 sample towers. The fusion method can be weighted average or comprehensive judgment based on certain specific rules. For example, if a risk knowledge vector has a higher weight or is more important in a specific evaluation indicator, it will be given a greater impact during fusion. Through this fusion, the first risk warning signal corresponding to each of the 10,000 sample towers is finally generated.

[0065] For the second model architecture: The server collaboratively updates the weight parameter information of the second model architecture to Z model processing units. Assume that the value of Z is 4.

[0066] At the same time, the server divides the 10,000 sample tower-related feature data into 4 sample feature data sequences. Each sequence contains at least one sample tower-related feature data, which may be divided according to certain rules, such as geographical location, data collection time, etc.

[0067] The first model processing unit is responsible for processing the first sample feature data sequence. For each sample tower associated feature data in the sequence, such as the first sample tower in the first sample feature data sequence, the first model processing unit predicts the corresponding second knowledge vector data based on the second model architecture to generate a corresponding second risk warning signal. Similarly, the second, third, and last sample towers in the first sample feature data sequence are processed in this way, thereby generating the second risk warning signal corresponding to each sample tower in the first sample feature data sequence.

[0068] Next, the server fuses the second risk warning signals corresponding to each sample tower in the four sample feature data sequences. The fusion method is similar to the fusion method in the first model architecture. By comprehensively considering the importance and weight of each signal, the second risk warning signals corresponding to each of the 10,000 sample towers are generated.

[0069] Through the above processing of the first model architecture and the second model architecture, the server can make full use of the structure and data characteristics of the model to perform collaborative training, thereby improving the performance and accuracy of the tower safety warning network model.

[0070] For example, in the actual processing process, for a specific sample tower, its first knowledge vector data may include the values ​​of multiple features such as the tower's height, tilt angle, material strength, etc. When processing, the first model processing unit will calculate and analyze these features according to the weight parameter information of the model block it loads, and obtain a preliminary risk assessment result, namely the first risk knowledge vector. The other four model processing units will also perform similar processing and generate their own corresponding risk knowledge vectors. Then, by fusing these risk knowledge vectors, a more comprehensive and accurate first risk warning signal is obtained.

[0071] For the second model architecture, it is assumed that the sample towers in the first sample feature data sequence mainly come from a specific area. When processing, the first model processing unit will comprehensively consider the various features and data relationships of these sample towers based on the second model architecture to generate a second risk warning signal corresponding to each sample tower. Finally, by fusing the risk warning signals in the four sample feature data sequences, a more comprehensive and reliable second risk warning signal is obtained.

[0072] In this way, through the first collaborative training mechanism, the server can fully explore and utilize the potential of the tower safety warning network model, improve the prediction and warning capabilities of the tower safety status, and ensure the stable operation of the communication network.

[0073] For another example, for the first model architecture, suppose one of the model blocks is mainly responsible for processing features related to the tower height. After the first model processing unit loads the weight parameter information of this block, it can perform targeted analysis and prediction on the height features in the input sample tower-related feature data, and generate a risk knowledge vector related to the height. Other model blocks and model processing units will also perform similar processing on the features they are responsible for, and finally obtain a comprehensive first risk warning signal through fusion.

[0074] For the second model architecture, it is assumed that the sample towers in the second sample feature data sequence are all under similar climatic conditions. When processing, the second model processing unit can fully consider the commonality of such climatic conditions, accurately predict the second knowledge vector data of each sample tower, and generate a corresponding second risk warning signal. Then, by fusing the signals of each sample feature data sequence, a comprehensive risk assessment result is obtained.

[0075] In actual operation, the server may continuously adjust the division of model blocks, the number of model processing units, and the fusion methods and strategies to adapt to different data sets and business needs, thereby continuously optimizing the performance and effect of the tower safety warning network model.

[0076] For example, considering that the monitoring data of the tower may fluctuate greatly in certain time periods, the server will adjust the weight distribution of the model blocks in the first model architecture accordingly, or re-divide the sample feature data sequence to ensure that the model can capture these changes more accurately. If it is found that there are more safety hazards in the towers in a certain area, the server can perform more detailed processing and analysis on the sample data in that area, adjust the parameters and architecture of the model, and improve the safety warning capability of the towers in that area.

[0077] Suppose at a certain moment, the server receives a large amount of monitoring data about the newly built tower. These new data are different from the previous data, such as the innovative design structure of the tower, or the special geographical environment. The server will immediately start the first collaborative training mechanism to adjust and optimize the first model architecture and the second model architecture accordingly. For the first model architecture, the model blocks may be re-divided to better adapt to the feature distribution of the new data. For the second model architecture, the sample feature data sequence may be re-divided, and the data of the newly built tower will be grouped into a separate sequence for key processing and analysis.

[0078] In addition, if the computing resources of the server change, such as adding a batch of high-performance computing nodes, the server can increase the number of model processing units, improve parallel processing capabilities, and speed up model training and prediction. Conversely, if computing resources are tight, the server may reduce the number of model processing units or simplify the model architecture to ensure that the basic functions of the model are not affected.

[0079] In actual business scenarios, some emergencies may occur, such as severe weather causing a large number of towers to face risks at the same time. At this time, the server needs to respond quickly and make urgent adjustments and optimizations to the tower safety warning network model. Through the first collaborative training mechanism, model blocks and sample feature data sequences can be quickly reallocated, and computing resources can be concentrated to process and analyze key data, and accurate safety warning results can be generated in a timely manner to provide decision support for maintenance personnel.

[0080] Assume that after a strong wind, the server receives tower monitoring data from multiple regions. Through the coordinated processing of the first model architecture and the second model architecture, the server can quickly screen out towers that are most affected by the strong wind and generate corresponding risk warning signals. For towers located in key wind disaster areas, the server will give higher weight and attention to ensure accurate assessment of their safety status.

[0081] Therefore, through the continuous optimization and adjustment of the tower safety early warning network model by the first collaborative training mechanism, timely, accurate and reliable early warning services can be provided for the safe operation of the tower in various complex scenarios.

[0082] In a possible implementation manner, the weight parameter information of the first model architecture includes weight parameter information corresponding to each of a plurality of risk knowledge vectors.

[0083] Step S131 includes: The weight parameter information corresponding to each of the plurality of risk knowledge vectors is predicted to generate the weight coefficients corresponding to each of the plurality of risk knowledge vectors.

[0084] The weight coefficients corresponding to the plurality of risk knowledge vectors are grouped and calculated with Y to generate group numbers corresponding to the plurality of risk knowledge vectors.

[0085] The weight parameter information of the risk knowledge vectors corresponding to the same grouping number is allocated to the same model block to generate the weight parameter information of the Y model blocks.

[0086] The weight parameter information of the Y model blocks is collaboratively updated to the Y model processing units.

[0087] In this embodiment, when processing the first model architecture, its weight parameter information includes weight parameter information corresponding to each of the plurality of risk knowledge vectors.

[0088] First, the server predicts the weight parameter information corresponding to each of these multiple risk knowledge vectors. It is assumed that these risk knowledge vectors include features related to the tower's height, tilt angle, material strength, wind speed impact, temperature change, etc.

[0089] For the risk knowledge vector related to tower height, the server generates a corresponding weight coefficient by analyzing historical data and performance in model training, taking into account the relative importance of tower height to the overall risk assessment. For example, if the change in tower height has a greater impact on safety risks, a higher weight coefficient may be generated.

[0090] For the risk knowledge vector related to the tilt angle, the server also evaluates its role in risk assessment based on data characteristics and model training, and generates corresponding weight coefficients. If a slight change in the tilt angle may cause a greater safety hazard, then its weight coefficient will be correspondingly higher.

[0091] The server performs such predictions on all risk knowledge vectors related to various features, thereby generating weight coefficients corresponding to multiple risk knowledge vectors.

[0092] Next, the server performs group calculations on these weight coefficients and Y. Assume that Y is set to 6.

[0093] Taking the weight coefficient of the highly correlated risk knowledge vector as an example, the server uses a specific algorithm to calculate it with 6 to obtain a group number. Similarly, the weight coefficients of the risk knowledge vectors related to the tilt angle, material strength, wind speed impact, temperature change, etc. are also grouped and calculated in this way to generate the group number corresponding to each risk knowledge vector.

[0094] Then, the server assigns the weight parameter information of the risk knowledge vectors corresponding to the same group number to the same model block. For example, the risk knowledge vector with group number 1, regardless of whether it is related to height, tilt angle or other features, has its weight parameter information assigned to model block 1. And so on, the weight parameter information of 6 model blocks is formed.

[0095] After the allocation is completed, the server collaboratively updates the weight parameter information of the six model blocks to the six model processing units. The first model processing unit loads the weight parameter information of model block 1, the second model processing unit loads the weight parameter information of model block 2, and so on.

[0096] In actual operation, suppose the server receives a batch of new tower monitoring data. For the first tower, the data of its height, tilt angle, material strength and other characteristics are input into the model.

[0097] The first model processing unit uses the weight parameter information of the loaded model block 1 to process and analyze the input feature data. For example, model block 1 contains weight parameter information related to height and some material strength. The model processing unit will calculate the input height and some material strength data based on these parameters to obtain a preliminary risk assessment result. The second model processing unit will also process other input feature data based on the weight parameter information of the loaded model block 2 to generate corresponding risk assessment results. By analogy, all six model processing units process and evaluate the input tower feature data. Afterwards, the server will combine the evaluation results of these six model processing units to generate a more comprehensive and accurate risk warning signal.

[0098] For example, in another scenario, the server found that towers in certain areas are more sensitive to the characteristics of tilt angle and foundation stability due to special geological conditions.

[0099] The server will re-predict and adjust the weight parameter information of these risk knowledge vectors. For risk knowledge vectors related to tilt angle and foundation stability, higher weight coefficients are generated to reflect their importance in these specific areas. Then, according to the group calculation and allocation process, the weight parameter information of the model block is readjusted to ensure that the model can more accurately warn of risks to towers in these areas.

[0100] Assume that after a period of operation, the server analyzes a large number of actual cases and finds that the impact of wind speed changes on tower safety is greater than previously expected. Therefore, the server re-predicts the weight parameter information of the risk knowledge vector related to wind speed and generates a higher weight coefficient. Then, through group calculation and redistribution, it is allocated to a more appropriate model block, and the weight parameter information of the model processing unit is updated to improve the model's emphasis on wind speed factors and the accuracy of risk assessment.

[0101] In this way, the first model architecture can always adapt to different situations and needs, and provide more reliable and accurate services for tower safety warning.

[0102] In a possible implementation, step S131 includes: Determine the risk knowledge labels corresponding to each weight parameter information of the first model architecture.

[0103] The weight parameter information corresponding to the same risk knowledge label is allocated to the same model block to generate the weight parameter information of the Y model blocks.

[0104] The weight parameter information of the Y model blocks is collaboratively updated to the Y model processing units.

[0105] In this embodiment, first, the server needs to determine the risk knowledge labels corresponding to each weight parameter information of the first model architecture. Assume that in this scenario, the risk knowledge labels faced by the server include "structural stability related", "environmental factor impact related", "material aging related", etc.

[0106] The server conducts an in-depth analysis of the weight parameter information in the first model architecture. For example, some weight parameter information is closely related to the basic structural strength of the tower, the inclination of the tower body, the connection tightness of the components, and other characteristics. The server marks it as a risk knowledge label related to "structural stability".

[0107] For other weight parameter information, such as those related to the calculation of the impact of environmental factors such as wind speed, temperature, and humidity on the tower, the server will give it a label of "related to environmental factors." Another part of the weight parameter information, which involves the evaluation of the fatigue and corrosion degree of the tower material, is marked as "related to material aging."

[0108] After determining these risk knowledge labels, the server starts to assign weight parameter information corresponding to the same risk knowledge label to the same model block.

[0109] Assume that the server assigns the weight parameter information corresponding to the label "related to structural stability" to model block 1. This block may contain specific weights for calculating the impact of the tower body inclination on risk, as well as parameters for evaluating the relationship between the tightness of component connections and safety risks. The weight parameter information corresponding to the label "related to the impact of environmental factors" is assigned to model block 2. This block may contain various weight parameters related to how environmental variables such as wind speed, temperature, and humidity affect the safety of the tower. The weight parameter information corresponding to the label "related to material aging" is assigned to model block 3.

[0110] Through this allocation method, the server successfully generates the weight parameter information of Y model blocks. Assume that Y is 5. In addition to the above three model blocks, there are also model blocks 4 and 5 corresponding to labels such as "maintenance history related" and "surrounding facility impact related".

[0111] Next, the server collaboratively updates the weight parameter information of the five model blocks to the five model processing units.

[0112] The first model processing unit receives and loads the weight parameter information of model block 1. In actual operation, when a new set of tower characteristic data is input, the first model processing unit can use these weight parameters related to structural stability to process and analyze the data. For example, for a newly monitored tower, its tower inclination data is input to the first model processing unit. The model processing unit calculates the risk assessment value of the inclination to the overall structural stability of the tower based on the loaded weight parameters.

[0113] The second model processing unit loads the weight parameter information of model block 2 and specifically handles calculations related to the impact of environmental factors. Assuming that the input environmental data shows that the current wind speed is extremely high and the temperature is extremely low, the model processing unit will evaluate the potential threat these environmental conditions pose to the safety of the tower based on the corresponding weight parameters.

[0114] The third model processing unit analyzes the aging of the tower material based on the weight parameters of model block 3. If a high degree of fatigue of the material is detected, a corresponding risk warning will be given.

[0115] Similarly, the 4th and 5th model processing units use the weight parameter information of model blocks 4 and 5 respectively to evaluate the maintenance history of the tower and the impact of surrounding facilities.

[0116] In the actual operation process, the server may encounter various specific situations.

[0117] For example, in a certain area, due to frequent extreme weather, the impact of environmental factors on tower safety becomes particularly important. The server will focus on the weight parameter information related to environmental factors in model block 2, and may fine-tune these parameters based on new data and actual conditions to ensure that the model can more accurately assess the risks brought by environmental factors.

[0118] Or, in a batch of newly built towers, new materials are used, and the material aging pattern is different from the past. The server will re-examine the weight parameters related to material aging in model block 3 and make necessary updates and adjustments.

[0119] For example, if the server finds that the maintenance work of a tower in a certain area is not standardized, which may affect the safety of the tower, the server will focus on checking the weight parameters related to the maintenance history in model block 4 to more accurately reflect the impact of the maintenance situation on the safety of the tower.

[0120] In addition, if a large building facility is newly added near a certain location, it may have an additional impact on the tower. The server will adjust the weight parameters related to the impact of surrounding facilities in model block 5 accordingly.

[0121] By continuously adjusting and optimizing the weight parameters according to actual conditions, and collaboratively updating the updated model block weight parameter information to the corresponding model processing unit, the server can ensure that the tower safety warning network model always maintains good performance and accuracy, providing strong support for ensuring the safe and stable operation of a large number of towers in the communication network.

[0122] Suppose that after a strong wind and rainstorm, the server received a large amount of monitoring data from the tower.

[0123] The first model processing unit uses the weight parameters of model block 1 to focus on analyzing the stability changes of the tower structure after severe weather, such as whether there is new tilt or loose connection of components.

[0124] The second model processing unit evaluates the direct impact and potential threats of extreme environmental conditions such as strong winds and rainstorms on the tower based on the weight parameters of model block 2, such as whether it exceeds the risk threshold set previously.

[0125] The third model processing unit uses the weight parameters of model block 3 to check the accelerated aging of the tower material after severe weather and determine whether maintenance or replacement is needed in advance.

[0126] The fourth model processing unit checks the past maintenance records of these towers based on the weight parameters of model block 4 to determine whether the previous maintenance measures are sufficient to cope with the current severe weather impact.

[0127] The fifth model processing unit uses the weight parameters of model block 5 to analyze whether the newly added buildings or other changes in the surrounding area have caused additional pressure and risks to the tower.

[0128] The server combines the analysis results of these five model processing units to generate a comprehensive and accurate tower safety warning report. If it is found that some towers have a high safety risk, the server will immediately send an alarm to the maintenance personnel to inform them to take urgent measures to repair and reinforce them.

[0129] In another scenario, the manufacturing technology and materials of the tower have been updated over time.

[0130] The server will redefine the risk knowledge labels of the weight parameters related to new materials and new technologies in the first model architecture. For example, for a new high-strength material, its related weight parameters may be labeled as "new material performance related".

[0131] Then, the weight parameter information with the same risk knowledge label is assigned to the new or corresponding model block.

[0132] The updated model block weight parameter information will be collaboratively updated to the corresponding model processing unit, so that the model can adapt to new tower technology and material changes and provide more accurate safety warnings.

[0133] Let's assume that in an area with complex geographical environment, the tower faces various risk factors, such as earthquake activity, landslides, etc.

[0134] The server will determine new risk knowledge labels for these special risk factors, such as "geological disaster impact related".

[0135] The relevant weight parameter information is allocated to a dedicated model block and updated to the corresponding model processing unit to enhance the model's assessment and early warning capabilities for special risks of towers in the area.

[0136] In short, by continuously determining risk knowledge labels, assigning weight parameters, and updating model processing units, the server can enable the tower safety warning network model to always adapt to various complex and changeable actual situations, providing reliable protection for the safe operation of the tower.

[0137] In a possible implementation manner, the weight parameter information of the first model architecture includes weight parameter information corresponding to each of a plurality of risk knowledge vectors.

[0138] Step S131 includes: For the weight parameter information corresponding to each of the multiple risk knowledge vectors and having a scale smaller than the first threshold value, the weight parameter information is segmented based on the risk knowledge label corresponding to the weight parameter information to generate the first weight parameter sub-information corresponding to each of the Y model blocks.

[0139] For the weight parameter information corresponding to each of the plurality of risk knowledge vectors, the weight coefficient of the weight parameter information is determined, the weight coefficient of the weight parameter information is grouped and calculated with Y to generate a group number of the weight parameter information, and the second weight parameter sub-information corresponding to each of the Y model blocks is generated based on the group number of the weight parameter information.

[0140] Based on the first weight parameter sub-information corresponding to each of the Y model blocks and the second weight parameter sub-information corresponding to each of the Y model blocks, weight parameter information of the Y model blocks is generated.

[0141] The weight parameter information of the Y model blocks is collaboratively updated to the Y model processing units.

[0142] In this embodiment, the weight parameter information of the first model architecture includes weight parameter information corresponding to each of the plurality of risk knowledge vectors.

[0143] First, the server needs to determine a first threshold value, assuming that the first threshold value is set to 1000. The server starts to classify the weight parameter information corresponding to each of the multiple risk knowledge vectors.

[0144] For weight parameter information whose scale is less than the first threshold value (1000), the server segments it based on its corresponding risk knowledge label. Assume that these risk knowledge labels include "related to minor structural changes", "related to short-term environmental fluctuations", "related to small-scale material wear", etc. For example, some weight parameter information related to minor structural changes of the tower (such as tiny welding cracks) has a small scale. The server reasonably segments these weight parameter information based on the risk knowledge label "related to minor structural changes" and generates the first weight parameter sub-information corresponding to each of the Y model blocks.

[0145] For weight parameter information whose scale is not less than the first threshold value (1000), the server first determines its weight coefficient. For example, if the scale of weight parameter information related to long-term and significant environmental changes (such as the impact of continuous strong winds) is large, the server calculates its weight coefficient.

[0146] Then, the server performs group calculation on the weight coefficient and Y. Assuming Y is set to 8, corresponding group numbers are generated for these weight parameter information through group calculation. Based on the group number, second weight parameter sub-information corresponding to each of the 8 model blocks is generated.

[0147] Next, the server generates weight parameter information of the eight model blocks by combining the first weight parameter sub-information and the second weight parameter sub-information generated previously.

[0148] For example, model block 1 may include first weight parameter sub-information related to "slight structural changes" and some second weight parameter sub-information of larger scale allocated by grouping number.

[0149] After completing the generation of the model block weight parameter information, the server collaboratively updates the weight parameter information of the eight model blocks to the eight model processing units.

[0150] The first model processing unit receives and loads the weight parameter information of model block 1. When new tower data is input, it can perform calculations and analyses based on these weight parameters.

[0151] Assume that at a certain moment, the server receives a set of tower monitoring data from a specific area.

[0152] The first model processing unit uses the weight parameter information it loads to evaluate the data related to slight structural changes, while also taking into account the impact of the larger-scale weight parameter information, and gives a preliminary risk analysis result.

[0153] The second model processing unit to the eighth model processing unit also perform corresponding processing and analysis according to the model block weight parameter information loaded therein.

[0154] During the actual operation process, the server will be adjusted and optimized according to different situations.

[0155] For example, in a certain season, due to frequent weather changes, the impact of short-term environmental fluctuations on tower safety becomes more important. The server will review the weight parameter information related to "short-term environmental fluctuations" and may adjust the segmentation method or group calculation rules to more accurately reflect its role in the model.

[0156] Alternatively, as the tower is used for a longer period of time, some originally small-scale risk factors (such as small-scale material wear) gradually develop into larger risks. The server will re-evaluate the scale of these weight parameter information and adjust it from segmented processing based on risk knowledge labels to processing through weight coefficients and group numbers to increase its weight and attention in the model.

[0157] In addition, if the communication network where the server is located is upgraded and the tower's load requirements and working environment change, the server will comprehensively re-evaluate all weight parameter information. The weight parameter information corresponding to the risk knowledge label that is no longer important may reduce its proportion in the model block; and for the newly emerging key risk factors, the scale and importance of their weight parameter information will be increased accordingly.

[0158] Suppose that after an extreme weather event, the server receives a large amount of data on damaged towers.

[0159] The first model processing unit focuses on processing weight parameter information related to minor structural changes, analyzing structural damage that may not be obvious but still poses a risk.

[0160] The second model processing unit focuses on processing weight parameter information related to short-term environmental fluctuations and assessing the short-term but strong impact of this extreme weather.

[0161] Similarly, the 3rd to 8th model processing units analyze and evaluate different aspects of the tower according to the weight parameter information in their respective model blocks.

[0162] The server combines the analysis results of these eight model processing units to generate a comprehensive and accurate tower safety assessment report. If it is found that some towers have a high safety risk, the server will immediately send an alarm to the maintenance personnel to arrange emergency maintenance and repair work.

[0163] In another scenario, thanks to technological advances, new monitoring equipment is able to provide more detailed data on the internal structure of the tower material.

[0164] The server will redefine the scale of the weight parameter information and risk knowledge labels related to the internal structure of the material based on these new data characteristics.

[0165] For weight parameter information with a smaller scale but increased importance, re-segmentation is performed; for weight parameter information with a larger scale, the weight coefficient and group number are recalculated.

[0166] Then the weight parameter information of the model block is updated and collaboratively updated to the corresponding model processing unit, so that the model can more accurately assess the safety risks of the tower due to changes in the internal structure of the material.

[0167] In short, by continuously classifying, processing, updating and collaboratively distributing weight parameter information, the server can ensure that the tower safety warning network model always adapts to various complex and changeable actual situations, and provides strong support for ensuring the safe and stable operation of a large number of towers in the communication network.

[0168] In a possible implementation, after step S135, the method further includes: Step A110, for the bth model processing unit, based on the second risk warning signals corresponding to each sample tower-associated characteristic data in the bth sample characteristic data sequence, adjust the weight parameter information of the bth model processing unit to generate iterative weight parameter information of the bth model processing unit.

[0169] Step A120: Synchronize the iteration weight parameter information corresponding to each of the Z model processing units to generate the iteration weight parameter information of the second model architecture.

[0170] Step A130: adjusting the weight parameter information of the Y model processing units based on the second risk warning signals corresponding to the X sample tower-related feature data, and generating iterative weight parameter information corresponding to the Y model processing units.

[0171] In this embodiment, it is assumed that the second model architecture divides a large amount of sample tower-related feature data into 5 sample feature data sequences, that is, Z is 5.

[0172] For the first model processing unit (b=1), it predicts the second knowledge vector data corresponding to each sample tower-associated feature data in the first sample feature data sequence based on the second model architecture, and successfully generates the second risk warning signal corresponding to each sample tower-associated feature data in the first sample feature data sequence.

[0173] Next, the server will adjust the weight parameter information for the first model processing unit. The server carefully analyzes the second risk warning signals corresponding to the associated feature data of each sample tower in the first sample feature data sequence. If these warning signals show that some prediction results deviate from the actual situation, or are not accurate and precise enough, the server will adjust the weight parameter information of the first model processing unit based on these feedbacks. For example, if the actual risk of some towers is high, but the second risk warning signal given by the first model processing unit is low, the server will increase the weight parameters related to these features accordingly, thereby generating iterative weight parameter information for the first model processing unit.

[0174] Likewise, the server performs the same operation on the second, third, fourth, and fifth model processing units.

[0175] After completing the adjustment of the weight parameters of each of the five model processing units, the server will synchronize the iterative weight parameter information corresponding to each of the five model processing units. The server will comprehensively consider the adjustment of each model processing unit to ensure that they maintain consistency and coordination as a whole, thereby generating the iterative weight parameter information of the second model architecture.

[0176] In addition, the server will also adjust the weight parameter information of the Y model processing units in the first model architecture based on the second risk warning signals corresponding to each of the X sample tower-related feature data. Assume that X is 10,000 and Y is 8.

[0177] The server will comprehensively analyze the second risk warning signals corresponding to each of the 10,000 sample tower-related feature data to identify the patterns and trends, as well as possible problems and deficiencies.

[0178] For example, if it is found that the overall risk warning effect is not ideal for certain types of towers (such as those located in high altitude areas or coastal areas), the server will specifically adjust the weight parameters related to the characteristics of these specific types of towers. Through such adjustments, the corresponding iterative weight parameter information is generated for the eight model processing units.

[0179] In actual operation, such a situation may occur. The server finds that the second risk warning signal of the towers in a certain area is generally not accurate due to the special geological conditions. For the model processing unit responsible for processing the sample feature data sequence of this area, the server will focus on adjusting its weight parameters and increase the weight of the features related to the geological conditions.

[0180] For another example, the server analyzes a large number of second risk warning signals and finds that there are deviations in the risk assessment of certain new tower structures. At this time, the server will simultaneously adjust the weight parameters of the relevant model processing units in the first model architecture and the second model architecture to improve the accuracy of risk warnings for such new tower structures.

[0181] Assume that, over a period of time, due to climate change, the wind load conditions faced by the tower have changed significantly. The server will adjust the weight parameters of the relevant model processing units according to the second risk warning signal corresponding to the new wind load data, so that the model can more accurately predict the tower safety risks caused by the change in wind load.

[0182] By continuously adjusting and optimizing the weight parameters of the model processing unit based on the second risk warning signal, the server can make the tower safety warning network model more accurate and reliable, and better ensure the safe and stable operation of a large number of towers in the communication network.

[0183] For example, after a large-scale storm, the server received the latest data of a large number of towers. By analyzing the second risk warning signals generated by each model processing unit, the server found that the risk warning for towers in areas severely affected by the storm was not timely and accurate enough.

[0184] In response to this situation, the server immediately adjusts the weight parameters of the model processing units responsible for the sample feature data sequences in these areas. For the model processing units in the first model architecture, the server may increase the weights of features related to storm intensity and duration; for the model processing units in the second model architecture, the server may adjust the weights related to the post-storm tower structure damage features.

[0185] Through such real-time adjustment and optimization, the next time similar extreme weather conditions occur, the tower safety warning network model will be able to give risk warning signals more quickly and accurately, providing strong support for timely maintenance measures.

[0186] For another example, with the expansion of the communication network, a number of new iron towers with special designs are built. After processing and analyzing the sample feature data of these new iron towers, the server adaptively adjusts the weight parameters of the model processing unit according to the generated second risk warning signal.

[0187] For the model processing units in the first model architecture, the weights of the features related to the design characteristics of the new tower may be reallocated; for the model processing units in the second model architecture, the weights related to various environmental factors and usage conditions may be readjusted based on the actual operating data of the new tower.

[0188] In this way, the tower safety early warning network model can continuously adapt to new tower types and changing environmental conditions, and always maintain good early warning performance.

[0189] In short, through the above series of weight parameter adjustment and optimization operations based on the second risk warning signal, the accuracy and reliability of the tower safety warning network model can be continuously improved, playing a key role in ensuring the safe and stable operation of a large number of towers in the communication network.

[0190] In a possible implementation, step S140 includes: Step S141, collaboratively update the weight parameter information of the tower safety warning network model to Q model processing units, and divide the T sample tower-associated feature data of the tower safety warning network model into Q sample feature data sequences, each sample feature data sequence includes at least one sample tower-associated feature data, the c-th model processing unit among the Q model processing units is used to process the c-th sample feature data sequence among the Q sample feature data sequences, Q is a positive integer, and c is a positive integer not greater than Q.

[0191] Step S142: for the c-th model processing unit, the c-th model processing unit is used to perform feature prediction on each sample tower-related feature data in the c-th sample feature data sequence according to the tower safety early warning network model, and a risk prediction signal corresponding to each sample tower-related feature data in the c-th sample feature data sequence is generated. Based on the risk prediction signal corresponding to each sample tower-related feature data in the c-th sample feature data sequence, a weight optimization direction value obtained by the c-th model processing unit is generated.

[0192] Step S143, fusing the weight optimization direction values ​​respectively determined by the Q model processing units to generate the weight optimization direction value of the tower safety early warning network model.

[0193] Step S144, updating the weight parameter information of the tower safety warning network model according to the weight optimization direction value of the tower safety warning network model, and generating the target tower safety warning network model.

[0194] In this embodiment, it is assumed that Q is set to 6, and the tower safety warning network model has 1000 sample tower-related feature data, that is, T is 1000.

[0195] First, the server collaboratively updates the weight parameter information of the tower safety warning network model to the six model processing units. At the same time, the server divides the 1,000 sample tower-related feature data into six sample feature data sequences. Each sample feature data sequence contains a varying number of sample tower-related feature data.

[0196] The first model processing unit is responsible for processing the first sample feature data sequence, the second model processing unit is responsible for processing the second sample feature data sequence, and so on, the sixth model processing unit is responsible for processing the sixth sample feature data sequence.

[0197] For the first model processing unit, it performs feature prediction on each sample tower-related feature data in the first sample feature data sequence according to the tower safety early warning network model.

[0198] Assume that the first sample feature data sequence contains 200 sample tower-related feature data, which include the tower's height, tilt angle, material strength, wind speed, temperature and other characteristics. The first model processing unit will perform in-depth analysis and calculation on each of the 200 sample tower-related feature data.

[0199] For example, for the first sample tower’s associated feature data, it will comprehensively consider all features and generate a preliminary risk prediction signal through complex operations and function calculations within the model. Then, the second sample tower’s associated feature data will be processed in the same way, and so on, until all 200 sample tower associated feature data in the first sample feature data sequence are processed, and a corresponding risk prediction signal is generated for each sample tower’s associated feature data.

[0200] Based on these risk prediction signals, the first model processing unit will further analyze and summarize them to generate the weight optimization direction value obtained by it. This weight optimization direction value reflects the direction and degree to which the first model processing unit believes that the weight parameters should be adjusted when processing this part of the data.

[0201] Similarly, the second to sixth model processing units will also perform similar processing on the sample feature data sequences they are responsible for, generate their own corresponding risk prediction signals, and generate their own weight optimization direction values ​​based on these risk prediction signals.

[0202] The server will merge the weight optimization direction values ​​determined by the six model processing units. During the fusion process, the server may use weighted average, median or other comprehensive evaluation methods to generate the weight optimization direction value of the tower safety warning network model.

[0203] This comprehensive weight optimization direction value takes into account the processing results and optimization suggestions of all model processing units and is a comprehensive and balanced direction guide.

[0204] According to the generated weight optimization direction value, the server updates the weight parameter information of the tower safety early warning network model. By adjusting the weight parameters, the model can better adapt to the data characteristics and prediction requirements, thereby generating the target tower safety early warning network model.

[0205] In actual operation, the following situation may occur: the server finds that in the sample feature data sequence processed by the first model processing unit, the height feature of most towers has a greater impact on risk prediction, but the current model does not allocate enough weight to this feature. Therefore, the weight optimization direction value generated by the first model processing unit will recommend increasing the weight parameter related to the tower height.

[0206] In the sample feature data sequence processed by the third model processing unit, the impact of changes in ambient temperature on tower safety is smaller than the model originally expected, so the weight optimization direction value generated by the third model processing unit will recommend reducing the weight parameters related to temperature.

[0207] When the server merges these six weight optimization direction values, it will take these different suggestions and situations into account. If multiple model processing units all suggest adjusting the weight parameters of a feature in the same direction, then this adjustment direction will be more clearly reflected in the merged weight optimization direction value.

[0208] For example, during an update process, the server finds that all four model processing units recommend increasing the weight parameters related to the tower material strength. In the final generated weight optimization direction value, the amplitude of increasing the material strength weight parameter will be relatively large.

[0209] For example, in some special cases, only individual model processing units may propose unique weight optimization directions. The server will carefully evaluate these suggestions and decide whether to adopt them and how to reflect them in the fusion based on the overall data distribution and model performance.

[0210] Assuming that the sample feature data sequence processed by the fifth model processing unit contains some tower data in special geographical locations, it recommends adjusting the weight parameters related to the geographical location. The server will carefully analyze the characteristics and impact of these special data, judge the rationality and universality of this suggestion, and then give appropriate consideration when integrating the weight optimization direction value.

[0211] By continuously performing such collaborative training and updating weight parameters, the server can make the tower safety warning network model more and more accurate and effective, and better provide reliable warning services for tower safety in communication networks.

[0212] Assume that, in a certain period of time, the tower data received by the server contains new feature information about the electromagnetic interference intensity around the tower. In the initial stage, the model may not process this feature accurately.

[0213] After analysis by each model processing unit and the generated weight optimization direction values, the server found that it was necessary to significantly increase the weight parameters of the electromagnetic interference intensity feature in the model.

[0214] Therefore, when the weight parameters are updated, the server will optimize the direction value according to the generated weight and adjust the weight of this feature accordingly. After multiple iterations and updates, the model can gradually and accurately process and utilize this new feature information, improving the ability to predict tower safety risks.

[0215] In short, through a series of operations based on the second collaborative training mechanism, the server can continuously optimize the tower safety warning network model so that it can adapt to the ever-changing tower data and security requirements and ensure the stable operation of the communication network.

[0216] In a possible implementation, step S140 may further include: Step S145, collaboratively update the weight parameter information of the tower safety warning network model to Q model processing units, and divide the W sample tower-associated feature data of the tower safety warning network model into U sample feature data sequences, each sample feature data sequence includes at least one sample tower-associated feature data, U is e times of Q, the d-th model processing unit among the Q model processing units is used to process e sample feature data sequences among the U sample feature data sequences, Q is a positive integer, d is a positive integer not greater than Q, U is a positive integer, and e is a positive integer.

[0217] Step S145: for the d-th model processing unit, the d-th model processing unit is used to perform feature prediction on each sample tower-related feature data in the e sample feature data sequences according to the tower safety early warning network model, and a risk prediction signal corresponding to each sample tower-related feature data in the e sample feature data sequences is generated. Based on the risk prediction signal corresponding to each sample tower-related feature data in the e sample feature data sequences, e sub-weight optimization direction values ​​obtained by the d-th model processing unit are generated.

[0218] Step S146: If the number of sample feature data sequences processed by the d-th model processing unit is less than the second threshold value, the e sub-weight optimization direction values ​​obtained by the d-th model processing unit are summarized to generate the weight optimization direction value obtained by the d-th model processing unit.

[0219] Step S147, if the number of sample feature data sequences processed by the dth model processing unit is not less than the second threshold value, the weight optimization direction values ​​respectively determined by the Q model processing units are merged to generate the weight optimization direction values ​​of the tower safety warning network model.

[0220] Step S148, based on the weight optimization direction value of the iron tower safety warning network model, the weight parameter information of the iron tower safety warning network model is updated to generate the target iron tower safety warning network model.

[0221] In this embodiment, it is further assumed that Q is set to 4, and the tower safety warning network model has 800 sample tower-related feature data, that is, W is 800. At the same time, U is set to 8, that is, each model processing unit needs to process 2 sample feature data sequences on average, that is, e is 2. In addition, the second threshold value is set to 3.

[0222] First, the server collaboratively updates the weight parameter information of the tower safety warning network model to the four model processing units. Then, the server divides the 800 sample tower-related feature data into 8 sample feature data sequences, each of which contains at least one sample tower-related feature data.

[0223] The first model processing unit is responsible for processing the first and second sample feature data sequences, the second model processing unit is responsible for processing the third and fourth sample feature data sequences, the third model processing unit is responsible for processing the fifth and sixth sample feature data sequences, and the fourth model processing unit is responsible for processing the seventh and eighth sample feature data sequences.

[0224] For the first model processing unit, it performs feature prediction on each sample tower-related feature data in the first and second sample feature data sequences respectively according to the tower safety early warning network model.

[0225] Assume that the first sample feature data sequence contains 120 sample tower-related feature data, and the second sample feature data sequence contains 80 sample tower-related feature data. For each sample tower-related feature data, such as the height, tilt angle, material strength, wind speed, temperature and other features of one of the towers, the first model processing unit will generate a corresponding risk prediction signal through model calculation and analysis.

[0226] After completing the processing of the two sample feature data sequences, based on the risk prediction signals corresponding to the respective sample tower-associated feature data, two sub-weight optimization direction values ​​obtained by the first model processing unit are generated, corresponding to the first and second sample feature data sequences respectively.

[0227] Since the number of sample feature data sequences processed by the first model processing unit is 2, which is less than the set second threshold value of 3, the server will aggregate the two sub-weight optimization direction values ​​to generate the weight optimization direction value obtained by the first model processing unit.

[0228] Similarly, the second model processing unit processes the third and fourth sample feature data sequences to generate corresponding risk prediction signals and two sub-weight optimization direction values. However, because the number of sample feature data sequences processed is 2, which is less than the second threshold value, it is also summarized to obtain the weight optimization direction value.

[0229] After the third model processing unit processes the fifth and sixth sample feature data sequences, the situation is different. Assuming that the fifth sample feature data sequence contains 150 samples and the sixth sample feature data sequence contains 130 samples, the number of sample feature data sequences processed is 2, which is also less than the second threshold value, and the weight optimization direction value is obtained by summarizing.

[0230] The fourth model processing unit processes the seventh and eighth sample feature data sequences, assuming that the seventh sequence has 100 samples and the eighth sequence has 120 samples. Since the number of sample feature data sequences processed by each model processing unit is less than the second threshold value, the server will not perform the fusion operation at this time.

[0231] However, if the number of sample feature data sequences processed by the fourth model processing unit is not less than the second threshold value, for example, the seventh sample feature data sequence contains 200 samples and the eighth sample feature data sequence contains 180 samples. Then, the server will fuse the weight optimization direction values ​​determined by the four model processing units to generate the weight optimization direction value of the tower safety warning network model.

[0232] Based on the generated weight optimization direction value, the server updates the weight parameter information of the tower safety warning network model. For example, if the weight optimization direction value indicates that the weight of certain features needs to be increased, the server will adjust the weight parameters of these features accordingly. Through such an update, the target tower safety warning network model is generated, which can more accurately predict and warn the safety status of the tower.

[0233] In actual operation, this may happen. In the first sample feature data sequence processed by the first model processing unit, most towers are located in mountainous areas, which are significantly affected by wind speed and terrain, but the model currently does not take these factors into consideration. Based on this, the generated sub-weight optimization direction value recommends increasing the weights of features related to wind speed and terrain.

[0234] In the third sample feature data sequence processed by the second model processing unit, the material aging problem of the tower is more prominent, but the model's prediction in this regard is not accurate enough. The corresponding sub-weight optimization direction value will suggest adjusting the weight parameters related to material aging.

[0235] In the fifth sample feature data sequence processed by the third model processing unit, it was found that the change in the tilt angle of some towers had a greater impact on the safety risk than the model expected. The generated sub-weight optimization direction value will propose to enhance the weight of the tilt angle feature in the model.

[0236] The seventh sample feature data sequence processed by the fourth model processing unit contains a large number of tower data from coastal areas, which are significantly affected by sea breeze erosion. The sub-weight optimization direction value generated by it will suggest increasing the weight of the features related to sea breeze erosion.

[0237] When the server aggregates or fuses these weight optimization direction values, it takes all situations into consideration. If multiple model processing units recommend increasing the weight of a feature, then this adjustment will be more strongly reflected in the final weight optimization direction value.

[0238] For example, assuming that the 1st, 3rd, and 4th model processing units all recommend increasing the weights of features related to wind speed, then in the fused weight optimization direction value, the magnitude of increasing the weight of the wind speed feature will be relatively large.

[0239] For example, in some special cases, only individual model processing units may make unique weight adjustment suggestions. The server will carefully evaluate these suggestions and combine the overall data distribution and model performance to decide how to reflect them in the final weight optimization direction value.

[0240] Assume that the sample feature data sequence processed by the second model processing unit includes some towers made of new materials, whose performance is different from that of traditional materials. The weight adjustment suggestions proposed by the second model processing unit are relatively unique. The server will carefully analyze the data characteristics and impacts of these new material towers, judge the rationality and universality of this suggestion, and then give appropriate consideration when integrating the weight optimization direction values.

[0241] By continuously performing such collaborative training and updating weight parameters, the server can make the tower safety warning network model more and more accurate and effective, and better provide reliable warning services for tower safety in communication networks.

[0242] Assume that, over a period of time, the tower maintenance strategy in the communication network has changed, resulting in a different distribution of some characteristic data of the tower. When the server processes according to the second collaborative training mechanism, it will generate corresponding weight optimization direction values ​​according to the new data situation and update the weight parameters of the model.

[0243] For example, the new maintenance strategy increases the frequency of material replacement for the tower, which affects the importance of material aging characteristics in risk assessment. Each model processing unit captures this change when processing the sample feature data sequence and reflects it in the generated sub-weight optimization direction value. The server updates the model weight parameters based on these weight optimization direction values ​​so that the model can adapt to the impact of the new maintenance strategy.

[0244] In short, through a series of operations based on the second collaborative training mechanism, the server can continuously optimize the tower safety warning network model so that it can adapt to the ever-changing tower data and operation and maintenance environment, and ensure the stable operation of the communication network.

[0245] In a possible implementation, the model knowledge features include: model architecture data of the tower safety warning network model.

[0246] The step of determining the model attribute label of the iron tower safety early warning network model based on the model knowledge feature includes: The scale of the tower safety early warning network model is calculated based on the model architecture data of the tower safety early warning network model.

[0247] If the scale of the iron tower safety early warning network model is not less than the third threshold value, the model attribute label of the iron tower safety early warning network model is marked as the first model attribute label.

[0248] If the scale of the iron tower safety early warning network model is less than the third threshold value, the model attribute label of the iron tower safety early warning network model is marked as the second model attribute label.

[0249] Alternatively, in another possible implementation, the model knowledge characteristics include: operating environment characteristics of the tower safety warning network model.

[0250] The step of determining the model attribute label of the iron tower safety early warning network model based on the model knowledge feature includes: If the operating environment characteristics of the tower safety warning network model belong to the first operating environment label, the model attribute label of the tower safety warning network model is marked as the first model attribute label.

[0251] If the operating environment characteristics of the tower safety warning network model belong to the second operating environment label, the model attribute label of the tower safety warning network model is marked as the second model attribute label, wherein the first operating environment label and the second operating environment label are different.

[0252] In this embodiment, first, the server obtains the model knowledge features of the tower safety warning network model, which includes the model architecture data. For example, the server obtains a tower safety warning network model, and its architecture data shows that it has 15 convolutional layers, and the number of filters in each convolutional layer starts from 32 and increases by a multiple of 1.5 until 256. There are 4 fully connected layers, and the number of neurons is 2048, 1024, 512 and 256 respectively.

[0253] Based on these model architecture data, the server begins to calculate the scale of the tower safety warning network model. When calculating the scale, the server will comprehensively consider factors such as the number of layers of the model, the number of neurons in each layer, the number of filters, and the connection method. For example, the server uses a specific algorithm to quantify and accumulate the number of neurons in each layer, the number of filters, and the connection weights between them. Assume that after calculation, the scale of this model is 40 million parameters.

[0254] At the same time, the server sets a third threshold value, assuming that the third threshold value is 35 million parameters. Since the calculated model scale of 40 million is greater than the third threshold value of 35 million, the server marks the model attribute label of the tower safety warning network model as the first model attribute label.

[0255] In another case, the architecture data of the tower safety warning network model obtained by the server shows that it has 8 convolutional layers, and the number of filters in each convolutional layer starts from 16 and increases in multiples of 2 until 128. There are 2 fully connected layers, with 512 and 256 neurons respectively. After calculating the scale, it is found that its scale is 28 million parameters. The set third threshold value is still 35 million parameters. Because the scale of the model of 28 million is less than the third threshold value of 35 million, the server marks the model attribute label of this tower safety warning network model as the second model attribute label.

[0256] In addition to determining the model attribute labels based on the model architecture data, the server also makes judgments based on the operating environment characteristics of the tower safety warning network model.

[0257] Assume that the server's operating environment has sufficient computing resources, including a large number of high-performance computing nodes, high-speed network connections, and large-capacity memory. This environment can easily handle large-scale model calculations and data processing tasks and is defined as the first operating environment label.

[0258] When the server obtains a tower safety warning network model and determines that its operating environment has the above-mentioned powerful computing resources and superior conditions, the server will mark the model attribute label of the tower safety warning network model as the first model attribute label.

[0259] On the contrary, if the server's operating environment has limited computing resources, low network bandwidth, and tight memory, it will be difficult to efficiently operate large-scale complex models. This environment is defined as the second operating environment label.

[0260] When the server obtains a tower safety warning network model and finds that it is in such an operating environment, the server will mark the model attribute tag of the tower safety warning network model as the second model attribute tag.

[0261] For example, in a server cluster in a certain area, the servers used to run the tower safety warning network model only have a few computing nodes with average performance, the network bandwidth is often limited, and the memory capacity is relatively small. When a new tower safety warning network model is connected, the server determines that it belongs to the second operating environment label after judging its operating environment characteristics, and thus marks the model attribute label of the model as the second model attribute label.

[0262] For example, in another large data center, the server has dozens of advanced high-performance computing servers, sufficient network bandwidth, and rich memory resources. When a new tower safety warning network model is deployed, the server marks the model attribute label of the model as the first model attribute label based on the powerful operating environment in which it is located.

[0263] Through the above judgment based on model architecture data and operating environment characteristics, the server can accurately label the corresponding model attribute labels for the tower safety warning network model, thereby providing an important basis for the subsequent selection of appropriate collaborative training mechanisms and network parameter learning methods to ensure the efficient operation and accurate prediction of the tower safety warning network model.

[0264] Therefore, by comprehensively considering the scale reflected by the model architecture data and the label type belonging to the operating environment characteristics, the attribute labels of the tower safety early warning network model can be accurately determined, and then the most appropriate training and optimization strategies can be adapted to improve the performance and effectiveness of the model in ensuring tower safety.

[0265] Figure 2 In an embodiment of the present application, a tower safety warning system 100 based on model analysis is provided, including a processor 1001 and a memory 1003 and an application code stored in the memory 1003. The processor 1001 executes the above application code to implement the steps of the sensor data processing method applied to smart photovoltaics.

[0266] Figure 2The tower safety warning system 100 based on model analysis shown includes: a processor 1001 and a memory 1003. The processor 1001 and the memory 1003 are connected, such as through a bus 1002. Optionally, the tower safety warning system 100 based on model analysis may also include a transceiver 1004, which may be used for data interaction between the server and other servers, such as data transmission and / or data reception. It should be noted that the transceiver 1004 is not limited to one in actual scheduling, and the structure of the tower safety warning system 100 based on model analysis does not constitute a limitation on the embodiments of the present application.

[0267] Processor 1001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. It may implement or execute various exemplary logic blocks, modules and circuits described in conjunction with the disclosure of this application. Processor 1001 may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0268] The bus 1002 may include a path to transmit information between the above components. The bus 1002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus. The bus 1002 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 2 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0269] The memory 1003 may be a ROM (Read Only Memory) or other types of static storage devices that can store static information and instructions, a RAM (Random Access Memory) or other types of dynamic storage devices that can store information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory) or other optical disk storage, optical disk storage (including compressed optical disk, laser disk, optical disk, digital versatile disk, Blu-ray disk, etc.), magnetic disk storage medium, other magnetic storage devices, or any other medium that can be used to carry or store application code and can be read by a computer, without limitation herein.

[0270] The memory 1003 is used to store application code for executing the embodiment of the present application, and the execution is controlled by the processor 1001. The processor 1001 is used to execute the application code stored in the memory 1003 to implement the steps shown in the above method embodiment.

[0271] The server includes but is not limited to: mobile terminals such as mobile phones, notebook computers, PADs, etc. and fixed terminals such as digital TVs, desktop computers, etc.

[0272] An embodiment of the present application provides a computer-readable storage medium on which application code is stored. When the application code is executed by a processor, the steps and corresponding contents of the aforementioned method embodiment can be implemented.

[0273] It should be understood that, although each operation step is indicated by arrows in the flowchart of the embodiment of the present application, the implementation order of these steps is not limited to the order indicated by the arrows. Unless there is clear explanation in this article, in some implementation scenarios of the embodiment of the present application, the implementation steps in each flowchart can be performed in other orders according to demand. In addition, some or all of the steps in each flowchart may include multiple sub-steps or multiple stages according to actual implementation scenarios, and some or all of these sub-steps or stages may be executed at the same time, and each sub-step or stage in these sub-steps or stages may also be executed at different times respectively. In different scenarios at the execution time, the execution order of these sub-steps or stages can be flexibly configured based on demand, and the embodiment of the present application does not limit this.

[0274] The above is only an optional implementation method for some implementation scenarios of the present application. It should be pointed out that for ordinary technicians in this technical field, without departing from the technical concept of the scheme of the present application, other similar implementation methods based on the technical ideas of the present application are also within the protection scope of the embodiments of the present application.

Claims

1. A tower safety early warning method based on model analysis, characterized in that: The method includes: Obtaining model knowledge features of a tower safety warning network model for which network parameter learning is to be performed; Based on the model knowledge characteristics, determining the model attribute label of the iron tower safety early warning network model; If the iron tower safety warning network model matches the first model attribute label, network parameter learning is performed on the iron tower safety warning network model according to a first collaborative training mechanism, wherein the first collaborative training mechanism is used to represent a composite training mechanism of a structure collaborative training mode and a data collaborative training mode; If the tower safety warning network model matches the second model attribute label, network parameter learning is performed on the tower safety warning network model according to the second collaborative training mechanism to generate a target tower safety warning network model, and the second collaborative training mechanism is used to represent a training mechanism of a data collaborative training mode; wherein the structure collaborative training mode is used to represent a training mode in which the model architecture of the tower safety warning network model is divided into multiple model processing units for network parameter learning, and the data collaborative training mode is used to represent a training mode in which the sample tower-related feature data of the tower safety warning network model is distributed to multiple model processing units for network parameter learning; The target tower safety warning network model is used to analyze the tower-related characteristic data of any input target tower to generate a corresponding safety warning result.

2. The tower safety early warning method based on model analysis according to claim 1 is characterized in that: The iron tower safety warning network model includes: a first model architecture based on the structure collaborative training mode, and a second model architecture based on the data collaborative training mode; wherein the model learning data of the first model architecture includes first knowledge vector data corresponding to each of the X sample iron tower associated feature data, the model prediction data of the first model architecture includes first risk warning signals corresponding to each of the X sample iron tower associated feature data, the model learning data of the second model architecture includes second knowledge vector data corresponding to each of the X sample iron tower associated feature data, the model prediction data of the second model architecture includes second risk warning signals corresponding to each of the X sample iron tower associated feature data, and X is a positive integer; The performing network parameter learning on the tower safety early warning network model according to the first collaborative training mechanism includes: For the first model architecture, the first model architecture is divided into Y model blocks, and the weight parameter information of the Y model blocks is collaboratively updated to Y model processing units, and the a-th model processing unit among the Y model processing units loads the weight parameter information of the a-th model block among the Y model blocks, where Y is a positive integer and a is not greater than Y; For the ath model processing unit, using the ath model block, feature prediction is performed on the first knowledge vector data corresponding to each of the X sample tower-related feature data, to generate the ath risk knowledge vector corresponding to each of the X sample tower-related feature data, where X is a positive integer; Fusing the Y risk knowledge vectors corresponding to the X sample tower-related feature data to generate first risk warning signals corresponding to the X sample tower-related feature data; For the second model architecture, the weight parameter information of the second model architecture is collaboratively updated to Z model processing units, and the X sample tower-related feature data are divided into Z sample feature data sequences, each sample feature data sequence includes at least one sample tower-related feature data, and the b-th model processing unit among the Z model processing units is used to process the b-th sample feature data sequence among the Z sample feature data sequences, where Z is a positive integer and b is a positive integer not greater than Z; For the b-th model processing unit, the b-th model processing unit is used to predict the second knowledge vector data corresponding to each sample tower-associated feature data in the b-th sample feature data sequence according to the second model architecture, and generate a second risk warning signal corresponding to each sample tower-associated feature data in the b-th sample feature data sequence; The second risk warning signals corresponding to each of the sample tower-associated feature data in the Z sample feature data sequences are fused to generate second risk warning signals corresponding to each of the X sample tower-associated feature data.

3. The tower safety early warning method based on model analysis according to claim 2 is characterized in that: The weight parameter information of the first model architecture includes weight parameter information corresponding to each of the plurality of risk knowledge vectors; The step of dividing the first model architecture into Y model blocks and collaboratively updating the weight parameter information of the Y model blocks to Y model processing units includes: Predicting weight parameter information corresponding to each of the plurality of risk knowledge vectors to generate weight coefficients corresponding to each of the plurality of risk knowledge vectors; Perform group calculations on the weight coefficients corresponding to the plurality of risk knowledge vectors and Y to generate group numbers corresponding to the plurality of risk knowledge vectors; Allocating the weight parameter information of the risk knowledge vectors corresponding to the same grouping number to the same model block to generate the weight parameter information of the Y model blocks; The weight parameter information of the Y model blocks is collaboratively updated to the Y model processing units.

4. The tower safety early warning method based on model analysis according to claim 2 is characterized in that: The step of dividing the first model architecture into Y model blocks and collaboratively updating the weight parameter information of the Y model blocks to Y model processing units includes: Determine the risk knowledge labels corresponding to the respective weight parameter information of the first model architecture; Allocating weight parameter information corresponding to the same risk knowledge label to the same model block to generate weight parameter information of the Y model blocks; The weight parameter information of the Y model blocks is collaboratively updated to the Y model processing units.

5. The tower safety early warning method based on model analysis according to claim 2 is characterized in that: The weight parameter information of the first model architecture includes weight parameter information corresponding to each of the plurality of risk knowledge vectors; The step of dividing the first model architecture into Y model blocks and collaboratively updating the weight parameter information of the Y model blocks to Y model processing units includes: For the weight parameter information whose scale is less than the first threshold value among the weight parameter information corresponding to each of the plurality of risk knowledge vectors, segment the weight parameter information based on the risk knowledge label corresponding to the weight parameter information to generate first weight parameter sub-information corresponding to each of the Y model blocks; For the weight parameter information corresponding to each of the plurality of risk knowledge vectors, the weight parameter information whose scale is not less than the first threshold value is determined; the weight coefficient of the weight parameter information is grouped and calculated with Y to generate a group number of the weight parameter information; based on the group number of the weight parameter information, the second weight parameter sub-information corresponding to each of the Y model blocks is generated; generating weight parameter information of the Y model blocks based on the first weight parameter sub-information corresponding to each of the Y model blocks and the second weight parameter sub-information corresponding to each of the Y model blocks; The weight parameter information of the Y model blocks is collaboratively updated to the Y model processing units.

6. The tower safety early warning method based on model analysis according to claim 2 is characterized in that: After predicting the second knowledge vector data corresponding to each sample tower-associated feature data in the b-th sample feature data sequence using the b-th model processing unit according to the second model architecture and generating the second risk warning signal corresponding to each sample tower-associated feature data in the b-th sample feature data sequence, the method further includes: For the b-th model processing unit, based on the second risk warning signals corresponding to the respective sample tower-associated feature data in the b-th sample feature data sequence, the weight parameter information of the b-th model processing unit is adjusted to generate iterative weight parameter information of the b-th model processing unit; Synchronizing the iteration weight parameter information corresponding to each of the Z model processing units to generate iteration weight parameter information of the second model architecture; The weight parameter information of the Y model processing units is adjusted based on the second risk warning signals corresponding to each of the X sample tower-related feature data, so as to generate iterative weight parameter information corresponding to each of the Y model processing units.

7. The tower safety early warning method based on model analysis according to claim 1 is characterized in that: The performing network parameter learning on the iron tower safety early warning network model according to the second collaborative training mechanism includes: The weight parameter information of the tower safety early warning network model is collaboratively updated to Q model processing units, and the T sample tower-related feature data of the tower safety early warning network model are divided into Q sample feature data sequences, each sample feature data sequence includes at least one sample tower-related feature data, and the c-th model processing unit among the Q model processing units is used to process the c-th sample feature data sequence among the Q sample feature data sequences, where Q is a positive integer and c is a positive integer not greater than Q; For the c-th model processing unit, the c-th model processing unit is used to perform feature prediction on each sample tower-related feature data in the c-th sample feature data sequence according to the tower safety early warning network model, and a risk prediction signal corresponding to each sample tower-related feature data in the c-th sample feature data sequence is generated; based on the risk prediction signal corresponding to each sample tower-related feature data in the c-th sample feature data sequence, a weight optimization direction value obtained by the c-th model processing unit is generated; The weight optimization direction values ​​respectively determined by the Q model processing units are merged to generate the weight optimization direction value of the tower safety early warning network model; According to the weight optimization direction value of the iron tower safety early warning network model, the weight parameter information of the iron tower safety early warning network model is updated to generate the target iron tower safety early warning network model.

8. The tower safety early warning method based on model analysis according to claim 1 is characterized in that: The performing network parameter learning on the iron tower safety early warning network model according to the second collaborative training mechanism includes: The weight parameter information of the tower safety early warning network model is collaboratively updated to Q model processing units, and the W sample tower-related feature data of the tower safety early warning network model are divided into U sample feature data sequences, each sample feature data sequence includes at least one sample tower-related feature data, U is e times Q, and the d-th model processing unit among the Q model processing units is used to process e sample feature data sequences among the U sample feature data sequences, Q is a positive integer, d is a positive integer not greater than Q, U is a positive integer, and e is a positive integer; For the d-th model processing unit, the d-th model processing unit is used to perform feature prediction on each sample tower-related feature data in the e sample feature data sequences according to the tower safety early warning network model, and a risk prediction signal corresponding to each sample tower-related feature data in the e sample feature data sequences is generated; based on the risk prediction signal corresponding to each sample tower-related feature data in the e sample feature data sequences, e sub-weight optimization direction values ​​obtained by the d-th model processing unit are generated; If the number of sample feature data sequences processed by the d-th model processing unit is less than the second threshold value, summarizing the e sub-weight optimization direction values ​​obtained by the d-th model processing unit to generate the weight optimization direction value obtained by the d-th model processing unit; If the number of sample feature data sequences processed by the d-th model processing unit is not less than the second threshold value, the weight optimization direction values ​​respectively determined by the Q model processing units are merged to generate the weight optimization direction value of the tower safety early warning network model; Based on the weight optimization direction value of the iron tower safety early warning network model, the weight parameter information of the iron tower safety early warning network model is updated to generate the target iron tower safety early warning network model.

9. The tower safety early warning method based on model analysis according to claim 1 is characterized in that: The model knowledge features include: model architecture data of the tower safety early warning network model; The step of determining the model attribute label of the iron tower safety early warning network model based on the model knowledge feature includes: Calculating the scale of the iron tower safety early warning network model based on the model architecture data of the iron tower safety early warning network model; If the scale of the iron tower safety early warning network model is not less than a third threshold value, marking the model attribute label of the iron tower safety early warning network model as the first model attribute label; If the scale of the iron tower safety early warning network model is less than the third threshold value, marking the model attribute label of the iron tower safety early warning network model as the second model attribute label; Alternatively, the model knowledge characteristics include: operating environment characteristics of the tower safety early warning network model; The step of determining the model attribute label of the iron tower safety early warning network model based on the model knowledge feature includes: If the operating environment characteristics of the iron tower safety early warning network model belong to the first operating environment label, the model attribute label of the iron tower safety early warning network model is marked as the first model attribute label; If the operating environment characteristics of the tower safety warning network model belong to the second operating environment label, the model attribute label of the tower safety warning network model is marked as the second model attribute label; wherein the first operating environment label and the second operating environment label are different.

10. A tower safety early warning system based on model analysis, characterized in that: The tower safety warning system based on model analysis includes a processor and a machine-readable storage medium, in which machine-executable instructions are stored. The machine-executable instructions are loaded and executed by the processor to implement the tower safety warning method based on model analysis of any one of claims 1-9.

Images