A method for locating and maintaining loose iron tower bolts based on deep learning and voiceprint analysis.

Abstract

This invention discloses a method for locating and maintaining loose tower bolts based on deep learning and acoustic signatures. The method includes: First, acquiring current acoustic signature data of the tower bolts using pre-set sensors. Then, processing the acoustic signature data using a pre-trained deep learning model to determine the current loose bolt location. Next, identifying the target tower bolt corresponding to the current loose location and acquiring its acoustic signature data over the entire monitoring cycle to calculate the structural damage risk coefficient of the tower structure where the target bolt is located. Finally, determining the maintenance strategy for the target tower bolt based on this structural damage risk coefficient. This design fully utilizes the automation and precision of deep learning technology, enabling rapid and accurate location of loose tower bolts, thus improving the efficiency and safety of tower maintenance.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a method for locating and maintaining loose bolts on iron towers based on deep learning and voiceprint analysis. Background Technology

[0002] In industries such as telecommunications and power, towers are an important component of infrastructure.

[0003] For example, Chinese patent CN116839883A discloses a method and device for diagnosing loose bolts on steel towers based on voiceprint recognition, relating to the field of fault diagnosis technology. The method involves acquiring real-time tower information of the target tower, adaptively determining the target acoustic signal generated by an acoustic wave generator based on the real-time tower information, transmitting the target acoustic signal to the target tower, and collecting the voiceprint signals at each bolt connection point of the target tower. For the voiceprint signal at each bolt connection point, the MFCC feature vector of the voiceprint signal is extracted, and the MFCC feature vector is weighted and dimensionality reduced to obtain the target feature vector. The target feature vector is input into a preset acoustic model to obtain the diagnostic identification result of whether the bolt is loose. An adaptive active sound generation system is adopted, and the sound content is automatically analyzed and adapted according to the tower model, size, and environment, greatly optimizing the efficiency of voiceprint analysis. Through a preset acoustic model, the entire tower's bolt tightness is efficiently scanned and analyzed to determine bolt looseness, improving detection accuracy and work efficiency.

[0004] For example, Chinese Patent Publication No. CN112786059A discloses a method and apparatus for voiceprint feature extraction based on artificial intelligence. This invention includes the following steps: collecting speech data and non-speech data to establish a sample database; retrieving an audio file from the sample database and processing it to obtain an audio frame sequence; performing a Fourier transform on each frame in the audio frame sequence to obtain the corresponding spectrogram information; extracting time-domain and frequency-domain information to obtain time-domain features and frequency-domain features; aggregating the time-domain and frequency-domain features to obtain aggregated features; embedding the aggregated features into vectors to obtain a voiceprint feature vector; inputting the voiceprint feature vector into a convolutional neural network model for training to obtain a voiceprint feature model; acquiring speech data to be recognized and preprocessing it; inputting the preprocessed speech feature data into the voiceprint feature model to obtain a speech feature vector. This invention improves the accuracy and efficiency of voiceprint feature extraction.

[0005] For example, Chinese patent CN111076960B discloses a voiceprint quality detection method based on artificial intelligence algorithms, including a voiceprint storage module and a labeling module in a PC. The implementation method includes the following steps: First, the voiceprint features of known air conditioner faults are labeled by the labeling module, and these data are divided into training set and test set according to a certain ratio; Second, a suitable artificial intelligence deep learning type is designed for the features of air conditioner voiceprints; Third, the model is trained using the labeled voiceprint data; Fourth, the empirical model is tested using the test set to determine its correctness; Fifth, the voiceprint to be detected is input into the model, and the model provides the detected voiceprint data; Sixth, the empirical model is continuously optimized. This invention can accurately identify various quality problems such as air conditioners under interference and variable conditions, covering the four main categories of air conditioner faults, and designs a reasonable software architecture, enabling the system to have good scalability and adaptive learning capabilities.

[0006] To ensure the normal operation and safe use of the tower, it is necessary to inspect and maintain it regularly. One key task is to check whether the tower bolts are loose.

[0007] Traditional inspection methods mainly rely on manual inspections, but this approach is inefficient and poses safety risks. Furthermore, since many iron towers are located in remote areas, manual inspections are even more difficult. The existing technologies mentioned above are relatively conventional, involving simple voiceprint feature extraction and recognition. One invention is primarily applied to the rather unique scenario of "diagnosing loose iron tower bolts," but the loosening of iron tower bolts is related to many conditions, making it impossible to pinpoint and maintain the location of the loose bolts.

[0008] Therefore, how to use modern technology to quickly, accurately, and automatically locate the loose bolt positions of iron towers is an urgent problem to be solved. Summary of the Invention

[0009] The purpose of this invention is to provide a method for locating and maintaining loose bolts on iron towers based on deep learning and voiceprint analysis.

[0010] In a first aspect, embodiments of the present invention provide a method for locating and maintaining the loose bolt position of a steel tower based on deep learning and voiceprint analysis, including: Acquire current feedback acoustic data for tower bolts based on preset sensors; The pre-trained tower bolt loosening location model is called to process the current feedback voiceprint data to obtain the current tower bolt loosening location; Identify the target tower bolt corresponding to the current location of the loose tower bolt; The feedback acoustic data of the target tower bolt during the overall monitoring cycle is obtained, and the structural damage risk coefficient of the tower structure where the target tower bolt is located is calculated. Based on the structural damage risk coefficient, a maintenance strategy for the target tower bolts is determined.

[0011] In a second aspect, embodiments of the present invention provide a server system, including a server, the server being used to execute the method described in the first aspect.

[0012] Compared to existing technologies, the beneficial effects of this invention include: The method for locating and maintaining loose tower bolts based on deep learning and acoustic signatures, as disclosed in this invention, includes: First, acquiring current feedback acoustic signature data of the tower bolts using a pre-set sensor. Then, processing the acoustic signature data using a pre-trained deep learning model to obtain the current loose position of the tower bolt. Next, identifying the target tower bolt corresponding to the current loose position and acquiring the feedback acoustic signature data of that target bolt throughout the overall monitoring cycle, thereby calculating the structural damage risk coefficient of the tower structure where the target bolt is located. Finally, based on the structural damage risk coefficient, determining the maintenance strategy for the target tower bolt. This design fully utilizes the automation and precision of deep learning technology, enabling rapid and accurate location of loose tower bolts, thus improving the efficiency and safety of tower maintenance. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 A flowchart illustrating the steps of a method for locating and maintaining loose iron tower bolts based on deep learning and voiceprint analysis, provided in an embodiment of the present invention. Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0016] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0017] In order to solve the technical problems mentioned in the background art Figure 1 This is a flowchart illustrating the method for locating and maintaining loose tower bolts based on deep learning and voiceprint analysis, as provided in this embodiment. The following is a detailed description of this method.

[0018] Step S201: Obtain the current feedback acoustic data of the tower bolts based on the preset sensor. Step S202: Call the pre-trained tower bolt loosening location model to process the current feedback voiceprint data to obtain the current tower bolt loosening location; Step S203: Determine the target tower bolt corresponding to the current loose tower bolt position; Step S204: Obtain the feedback acoustic data of the target tower bolt during the overall monitoring cycle, and calculate the structural damage risk coefficient of the tower structure where the target tower bolt is located. Step S205: Based on the structural damage risk coefficient, determine the maintenance strategy for the target tower bolts.

[0019] In this embodiment of the invention, for example, acoustic fingerprint sensors are pre-installed at several key parts of the tower, such as bolt connections. These sensors can capture the minute sound fluctuations generated by the tower bolts when affected by external factors such as wind and temperature changes, i.e., acoustic fingerprint data. The server receives and stores the acoustic fingerprint data from these sensors in real time through a data acquisition system. The server continuously receives data streams from various sensors, which contain information such as the sound frequency and amplitude of the tower bolts at different times. The server stores this raw data for subsequent analysis. The server has a built-in deep learning model, which has been trained with a large amount of acoustic fingerprint data and corresponding bolt location labels, specifically designed to identify specific acoustic fingerprint patterns when bolts are loose. When new acoustic fingerprint data is received by the server, the server calls this model to process the data. The server inputs the newly collected acoustic fingerprint data into the pre-trained deep learning model. The model processes and analyzes this data, and by comparing it with known acoustic fingerprint patterns, identifies the bolt loosening location indicated by the current acoustic fingerprint data. The server knows the exact location of every bolt on the tower, so when the model identifies a loose bolt, the server can quickly pinpoint which bolt(s) are causing the problem. Based on the loose location information output by the model, the server searches its database for the corresponding bolt record, thus identifying the specific target tower bolt. The server not only stores the current acoustic signature data but also saves historical acoustic signature data for each bolt over a past period (monitoring cycle). By analyzing this historical data, the server can assess the long-term impact of bolt loosening on the tower structure. The server retrieves all acoustic signature data for the target tower bolt within the monitoring cycle and uses this data to calculate the damage risk coefficient caused by bolt loosening to the tower structure. This coefficient may consider factors such as the frequency and extent of bolt loosening, and its impact on surrounding structures. Once the server obtains the structural damage risk coefficient, it can formulate a specific maintenance strategy for the target tower bolt based on this coefficient and preset maintenance rules. The server compares the calculated structural damage risk coefficient with thresholds in the maintenance rule base and selects appropriate maintenance measures based on the comparison results, such as immediately dispatching personnel to inspect and tighten the bolts, arranging a regular inspection schedule, or conducting more in-depth structural analysis. The server sends these maintenance instructions to the relevant maintenance team to ensure the safety and stability of the tower.

[0020] In this embodiment of the invention, the tower bolt loosening location model is obtained through the following method.

[0021] (1) Obtain an abnormal soundprint instance set; the abnormal soundprint instance set includes multiple sets of first abnormal soundprint instances and multiple sets of second abnormal soundprint instances, wherein each set of first abnormal soundprint instances includes feedback soundprint data of the sample tower bolt at a preset loose position, and each set of second abnormal soundprint instances includes other feedback soundprint data of the sample tower bolt at other positions. (2) Load the first abnormal voiceprint instance into the tower bolt loosening location model and execute the training process to obtain the first learned cost parameter; (3) Load the second abnormal soundprint instance into the tower bolt loosening location model, and infer the sample tower bolt loosening location in the second abnormal soundprint instance through the tower bolt loosening location model to obtain a first inference result; the first inference result includes the first confidence value of each preset loosening location inferred by the tower bolt loosening location model as the sample tower bolt has become loose; (4) Based on the first confidence value and the location confidence value threshold, the learned second cost parameter is obtained; (5) Based on the first cost parameter and the second cost parameter, the model parameters of the tower bolt loosening position positioning model are optimized and adjusted to obtain the tower bolt loosening position positioning model after training.

[0022] In this embodiment of the invention, exemplarily, the server retrieves an abnormal soundprint instance set from the database. This set contains two different sets of abnormal soundprint instances. The first set consists of feedback soundprint data when sample tower bolts loosen at a preset loosening location, recording specific sound patterns when the bolts loosen. The second set consists of feedback soundprint data when sample tower bolts loosen at other locations, representing atypical patterns of bolt loosening sounds. The server inputs the first set of abnormal soundprint instances (i.e., feedback soundprint data of sample tower bolts at the preset loosening location) into the tower bolt loosening location localization model, initiating the model's training process. This process involves the model analyzing and learning the soundprint data to identify the correlation between the bolt loosening location and the soundprint pattern. The server also inputs the second set of abnormal soundprint instances (i.e., feedback soundprint data of sample tower bolts at other locations) into the tower bolt loosening location localization model, but this time to test the model's inference ability. The model needs to process this atypical soundprint data and attempt to infer the bolt loosening location. The server compares the confidence value in the first inference result output by the model with a preset location confidence value threshold. This threshold is used to determine the reliability of the model's inference results. If the reliability value at a certain location is higher than the threshold, the model's inference for that location is considered reliable; otherwise, it is considered unreliable. The server now has two sets of cost parameters: the first set is the cost parameters when the model learns the bolt loosening location from a preset loosening point, and the second set is the cost parameters when the model processes atypical voiceprint data and infers the bolt loosening location. The server will use these two sets of parameters to optimize and adjust the model's internal parameters to improve its performance and accuracy.

[0023] To more clearly describe the solutions provided in the embodiments of this application, a more detailed explanation is provided below.

[0024] In this embodiment of the invention, exemplarily, the abnormal soundprint instance set is a collection containing multiple instances of soundprint data. These instances are marked as "abnormal," meaning they represent the sound patterns generated by tower bolts in abnormal or loose states. Assume there is a database storing a large amount of soundprint data on tower bolts in different states. From this database, data exhibiting abnormal soundprint characteristics, such as the unique sound frequency and amplitude patterns generated when bolts are loose, can be selected to form an abnormal soundprint instance set. The first abnormal soundprint instance refers to the soundprint data instances collected when tower bolts loosen at a preset loosening location. These instances are representative because they are directly related to the sound characteristics of bolts loosening at a specific location. For example, a soundprint sensor is installed at a critical connection point (preset loosening location) on the tower. When the bolt at this location loosens, the sensor captures specific sound fluctuations and records this data. This data constitutes the first abnormal soundprint instance. The preset loosening location refers to some pre-determined critical parts on the tower structure where bolts are prone to loosening due to stress, wear, or other reasons. Therefore, these locations receive special attention, and acoustic signature sensors are installed for real-time monitoring. During the design and construction of the tower, engineers may identify certain bolt connection points as key monitoring targets based on experience or structural analysis. For example, areas subjected to high winds or locations where loosening has historically occurred are marked as preset loosening locations. Secondary anomalous acoustic signature instances refer to acoustic signature data instances collected when bolts loosen at non-preset loosening locations (i.e., other locations). While these instances are not as directly linked to loosening at a specific location as the primary anomalous acoustic signature instances, they still provide valuable information about bolt loosening sound patterns. Besides preset loosening locations, there are many other bolt connection points on the tower. When bolts in these non-critical locations loosen, although the impact on the overall structure may be small, acoustic signature sensors can still capture the sound fluctuations caused by this loosening. This data constitutes the secondary anomalous acoustic signature instances. Feedback acoustic signature data refers to data collected from acoustic signature sensors that reflects the sound characteristics of tower bolts at different times and under different conditions. This data, including sound frequency, amplitude, and duration, is crucial for analyzing and assessing bolt loosening. The acoustic sensor continuously monitors changes in the sound of the tower bolts and transmits this data to a server for analysis in real time. By processing and analyzing this feedback acoustic data, the server can identify the location and extent of bolt loosening, as well as its potential impact on the tower structure.

[0025] Furthermore, the tower bolt loosening location model is a machine learning model whose main task is to predict or locate the loose bolts on the tower by analyzing acoustic data. During training, this model learns the relationship between the characteristics of the bolt loosening sound and the loosening location. Assume a neural network model is built using deep learning techniques. During the training phase, the first collected abnormal acoustic instances are used as input data, along with the corresponding loosening location information as labels. The model gradually establishes a mapping relationship between sound features and loosening locations by learning patterns in this data. In the training process of the tower bolt loosening location model, the first abnormal acoustic instances are first divided into multiple batches, each containing a certain number of samples. Then, these batches of data are sequentially input into the model. After each input, the difference between the model's prediction result and the true label (i.e., the loss function value) is calculated, and the model's internal parameters are adjusted based on this difference to reduce prediction error. This process is repeated until the model's predictive ability reaches a satisfactory level. The learned first cost parameters refer to the model parameters adjusted by the optimization algorithm during model training; these parameters reflect the knowledge learned by the model from the first abnormal acoustic instances. Cost parameters are typically related to the model's loss function and represent the model's performance on the training data. In neural network models, the first learned cost parameters might include the weights and biases of neurons. These parameters are updated during training using optimization algorithms such as gradient descent to minimize the error generated by the model when predicting the first abnormal voiceprint instance. The final set of parameters represents the knowledge the model has learned from this data, used to identify similar bolt loosening situations in subsequent inference processes.

[0026] Furthermore, in this embodiment of the invention, inference refers to the process by which the tower bolt loosening location model receives second abnormal acoustic fingerprint instances as input and attempts to predict the possible locations where the bolts may have become loose based on the acoustic characteristics of these data. This prediction process is based on the knowledge and patterns learned by the model during the training phase. The first inference result refers to the prediction result obtained by the tower bolt loosening location model after inferring from the second abnormal acoustic fingerprint instances. This result includes information about the preset locations where the model believes the bolts may have become loose, as well as the confidence level of each predicted location. Suppose the model infers from a set of second abnormal acoustic fingerprint instances and obtains a prediction result containing three preset loosening locations, each location having a corresponding confidence value. This prediction result is called the first inference result. The higher the confidence value, the more confident the model is in believing that the location has become loose; the lower the confidence value, the less certain the model is about the prediction result for that location. The first confidence value refers to the prediction confidence level given by the tower bolt loosening location model for each preset loosening location in the first inference result. This value is usually a probability value between 0 and 1, used to quantify the model's confidence level regarding the probability of loosening at a specific location. In the first inference result, the model calculates a confidence value for each preset loosening location. For example, for location A, the model gives a confidence value of 0.8, meaning the model has 80% confidence that loosening has occurred at that location; for locations B and C, the model gives confidence values ​​of 0.5 and 0.3 respectively, indicating that the model's prediction results for these locations are relatively uncertain. These confidence values ​​constitute the first confidence value for each preset loosening location in the first inference result. The location confidence value threshold is a preset standard used to judge whether the model's inference of the loosening location is sufficiently reliable. When the first confidence value given by the model is higher than this threshold, the model's inference result is considered reliable; otherwise, the result is considered unreliable. Suppose the location confidence value threshold is set to 0.6. For the first confidence value of location A in the above example, 0.7, since it is higher than the threshold of 0.6, the model's inference of loosening at location A is considered reliable. The learned second cost parameter refers to the parameter obtained when further adjusting or optimizing the tower bolt loosening location model based on the first confidence value and the location confidence value threshold. These parameters may include model weights, biases, etc., reflecting the model's learning outcomes when processing second anomalous voiceprint instances. In practical applications, the model parameters can be adjusted based on the comparison between the first confidence value and the localization confidence value threshold. For example, for inferences where the first confidence value is below the threshold, the number of samples at these locations can be increased or the model's internal structure adjusted to improve its ability to recognize these locations. After such adjustments and optimizations, the resulting model parameters are the learned second cost parameters.

[0027] Furthermore, if a Convolutional Neural Network (CNN) from deep learning is used to build a model for locating loose bolts on iron towers, the model parameters will include the weights and biases of multiple convolutional layers, as well as parameters of other layers such as pooling layers and fully connected layers. These parameters are iteratively updated during training using optimization algorithms to enable the model to better fit the training data. Optimization refers to iteratively updating the model's parameters using optimization algorithms (such as gradient descent, stochastic gradient descent, etc.) to minimize the loss function value during training. This process aims to improve the model's ability to fit the training data while maintaining its generalization performance on unseen data. During the training of the model for locating loose bolts on iron towers, a loss function is defined to measure the difference between the model's predictions and the true labels. The optimization algorithm adjusts the model's parameters based on the value of this loss function (i.e., performs optimization), gradually reducing the loss function value. This allows the model to provide more accurate predictions when processing similar data. A successfully trained model for locating loose bolts on iron towers is one whose performance has stabilized after sufficient iterative optimization and adjustment, and which demonstrates good predictive ability on unseen data. This model has learned all the necessary knowledge to identify the location of loose bolts from voiceprint data. Assuming multiple rounds of iterative training, a model is obtained that can accurately predict the location of loose bolts on the test set. This model is the completed tower bolt loosening location model, which can be deployed in a real-world environment for real-time monitoring and analysis of tower bolt loosening.

[0028] In this embodiment of the invention, the aforementioned abnormal voiceprint instance set also includes a target variable corresponding to the first abnormal voiceprint instance. The target variable is used to indicate the true value of the sample tower bolt loosening position in the first abnormal voiceprint instance. The aforementioned step of loading the first abnormal voiceprint instance into the tower bolt loosening position localization model to perform the training process and obtain the learned first cost parameter can be implemented through the following example.

[0029] (1) Load the first abnormal soundprint instance into the tower bolt loosening location model, and infer the sample tower bolt loosening location in the first abnormal soundprint instance through the tower bolt loosening location model to obtain a second inference result; the second inference result includes the second confidence value of each preset loosening location inferred by the tower bolt loosening location model; (2) Based on the target variable and the second inference result, the learned first cost parameter is obtained.

[0030] In this embodiment of the invention, exemplarily, the server receives acoustic signature data from various tower sensors, which is categorized into an abnormal acoustic signature instance set. Within this set, there is a special type of instance, the first abnormal acoustic signature instance, which includes not only acoustic signature data but also target variables. These target variables are determined by experts based on on-site inspections or other reliable methods and are used to mark the actual loosening locations of sample tower bolts. Now, the server needs to use this data to train a tower bolt loosening location localization model. The server first loads the first abnormal acoustic signature instances into the tower bolt loosening location localization model. These instances contain acoustic signature data generated when bolts loosen at different locations, along with the corresponding target variables, i.e., the ground truth value of the actual loosening location. After loading the data, the server processes this acoustic signature data using the tower bolt loosening location localization model. The model analyzes the features in the acoustic signature data and attempts to infer the possible locations where the bolts may loosen. This inference process produces a second inference result, which includes the model's assessment of the loosening probability for each preset loosening location, i.e., a second confidence value. For example, the model might give a confidence value of 0.8 for bolt loosening at location A, 0.5 at location B, and 0.2 at location C. These confidence values ​​reflect the model's confidence level regarding the likelihood of loosening at each location. Next, the server uses the target variable (the true value of the actual loosening location) and the second inference result (the loosening location inferred by the model and its confidence value) to calculate the model's error in this inference. This error reflects the difference between the model's prediction and the actual situation. For example, if the target variable indicates that the bolt at location A is indeed loose, and the model gives a high confidence value (such as 0.8), then the model's prediction at this location is accurate, with a small error. Conversely, if the model gives a high confidence value at location B or location C, but the bolts at these locations are not actually loose, then the model's error will be large. Based on the calculated error, the server uses optimization algorithms to adjust the model's parameters to reduce errors in future inferences. This process is the model's learning process; through this process, the model gradually learns how to accurately identify the loosening location of bolts from voiceprint data. After multiple rounds of iterative training, when the model's performance stabilizes and it demonstrates good predictive ability on unseen data, the first learned cost parameters are obtained. These parameters reflect the knowledge and experience the model has learned from the first abnormal voiceprint instance. The server continuously monitors the model's performance and adjusts the model's parameters or structure as needed to further optimize its performance. This may include adding more training data, adjusting the model's internal structure, using more complex optimization algorithms, etc. Through continuous optimization and adjustment, the server can ensure that the tower bolt loosening location model remains in optimal condition in practical applications.

[0031] In this embodiment of the invention, the tower bolt loosening location model includes a first intermediate network structure and a second intermediate network structure; the aforementioned step of loading the first abnormal soundprint instance into the tower bolt loosening location model, and inferring the sample tower bolt loosening location in the first abnormal soundprint instance through the tower bolt loosening location model to obtain a second inference result can be implemented through the following example.

[0032] (1) Load the first abnormal voiceprint instance into the first intermediate network structure, and perform feature extraction operation on the first abnormal voiceprint instance through the first intermediate network structure to obtain the first feature vector; (2) Load the first feature vector into the second intermediate network structure, and perform scalar multiplication on the first feature vector using the category deviation parameter through the second intermediate network structure to obtain the first element sequence; wherein, the category deviation parameter includes multiple category deviation vectors, each category deviation vector corresponds to a preset loosening position, and the number of elements in the first element sequence is consistent with the number of preset loosening positions; (3) The first element sequence is normalized by the localization activation function to obtain the second element sequence, and the second element sequence is determined as the second inference result; wherein, the elements in the second element sequence represent the second confidence value of the sample iron tower bolt corresponding to the preset loose position inference of the iron tower bolt by the iron tower bolt loose position localization model.

[0033] In this embodiment of the invention, for example, the server has established a model for locating the loose bolt position of a steel tower. This model includes a first intermediate network structure and a second intermediate network structure. Now, the server will use a first abnormal acoustic fingerprint instance to train this model and demonstrate in detail how inference is made through these two network structures to finally obtain a second inference result. The server first loads the first abnormal acoustic fingerprint instance into the first intermediate network structure. This network structure is responsible for extracting useful features from the acoustic fingerprint data. Through a series of complex mathematical operations and transformations, the first intermediate network structure transforms the original acoustic fingerprint data into a more easily processed and analyzed form, namely, a first feature vector. For example, the first intermediate network structure may include multiple convolutional layers, pooling layers, and fully connected layers, which work together to extract features related to the bolt loosening position from the acoustic fingerprint data, such as frequency, amplitude, time-domain and frequency-domain characteristics. These features are encoded in the first feature vector, preparing for the next processing step. Next, the server loads the first feature vector into the second intermediate network structure. This network structure is responsible for inferring the possible location of bolt loosening based on the extracted features. To achieve this goal, the second intermediate network structure uses a set of class bias parameters to perform scalar multiplication on the first feature vector. These class bias parameters are a set of predefined vectors, each corresponding to a preset loosening location. When the first feature vector is scalar multiplied with these vectors, the result highlights features associated with specific loosening locations, thus helping the model more accurately infer the actual loosening location of the bolt. The result of the scalar multiplication is a sequence of first elements, with the number of elements matching the number of preset loosening locations. Each element represents the model's assessment of the probability of loosening at a specific loosening location. To transform the sequence of first elements into a more interpretable and comparable form, the server uses a localization activation function to standardize it. This activation function maps each element in the sequence to a specific range (e.g., between 0 and 1) to more intuitively represent the model's confidence level for each loosening location. After standardization, the server obtains a second sequence of elements, i.e., the second inference result. Each element in this sequence represents a second confidence value for the model's inference that the sample tower bolt is loose at the corresponding preset loosening location. These confidence values ​​can be used to evaluate the model's predictive accuracy and serve as a basis for subsequent optimization and adjustment. Through the above steps, the server successfully inferred the first abnormal soundprint instance using the tower bolt loosening location model and obtained a second inference result regarding the bolt loosening location. This result will provide valuable information and feedback for subsequent training and optimization.

[0034] In this embodiment of the invention, the step of obtaining the learned first cost parameter based on the target variable and the second inference result can be implemented through the following example.

[0035] (1) Based on the true value of the loose position in the target variable, obtain the corresponding first category representation vector from the category deviation vector, and determine the category deviation vector other than the first category representation vector in the category deviation vector as the second category representation vector; (2) Obtain the first similarity score between the first feature vector and the first category representation vector, and obtain the second similarity score between the first feature vector and each of the second category representation vectors; (3) Determine the deviation between the first similarity score and the preset similarity score threshold to obtain the first deviation value; (4) Based on the first deviation value and each of the second similarity scores, the learned first cost parameter is obtained; The first deviation value is inversely proportional to the first cost parameter, and the second similarity score is directly proportional to the first cost parameter.

[0036] In this embodiment of the invention, for example, the server has already inferred a first abnormal soundprint instance using a tower bolt loosening location model, obtaining a second inference result. Now, the server will combine the target variable (i.e., the ground truth of the actual loosening location) and the second inference result to calculate a first cost parameter, which will be used for subsequent optimization and adjustment of the model. First, based on the ground truth of the loosening location in the target variable, the server finds the corresponding first category representation vector from the category bias vector set. This vector represents the representation of the actual loosening location in the model. Simultaneously, the server labels other vectors in the category bias vector set besides the first category representation vector as second category representation vectors, which represent the model representations of other non-loosening locations or incorrectly predicted locations. Next, the server calculates the similarity between the first feature vector (features extracted from the first abnormal soundprint instance) and the first category representation vector, obtaining a first similarity score. This score reflects the model's prediction accuracy at the actual loosening location. Simultaneously, the server also calculates the similarity between the first feature vector and each of the second category representation vectors, obtaining a series of second similarity scores. These scores reflect the model's prediction error at other non-loosening locations. The server compares the first similarity score with a preset similarity score threshold to determine the deviation between them, i.e., the first deviation value. This deviation value reflects the degree of error in the model's prediction of the actual loosening location. Ideally, if the model's prediction is completely accurate, the first similarity score should be close to or equal to the preset similarity score threshold, in which case the first deviation value is small; conversely, if the prediction error is large, the first deviation value will also increase accordingly. Finally, the server calculates the first cost parameter based on the first deviation value and each second similarity score. This parameter integrates the model's prediction error at the actual loosening location and prediction error information at other non-loosening locations, used to evaluate the overall performance of the model and guide subsequent optimization adjustments. Specifically, the first cost parameter is inversely proportional to the first deviation value, i.e., the smaller the deviation value (the more accurate the prediction), the smaller the cost parameter; and directly proportional to the second similarity score, i.e., the larger the prediction error at other non-loosening locations (the higher the similarity score), the larger the cost parameter. Through the above steps, the server successfully calculates the learned first cost parameter by combining the target variable and the second inference results. This parameter will provide important feedback and guidance for the subsequent training and optimization of the model.

[0037] In the embodiments of the present invention, the step of performing a normalization operation on the first element sequence by locating an activation function to obtain a second element sequence can be implemented through the following example.

[0038] (1) Obtain the positioning adjustment parameters; the value of the positioning adjustment parameters shall not be less than 1; (2) The first element sequence is tuned using the positioning adjustment parameters to obtain the third element sequence; (3) The third element sequence is normalized by the localization activation function to obtain the second element sequence.

[0039] In this embodiment of the invention, for example, during the process of locating the loose bolt position of a steel tower by the server, after obtaining the first element sequence through the second intermediate network structure, the first element sequence needs to be standardized using a localization activation function. In some cases, to improve the model's localization accuracy and stability, the server will first perform optimization on the first element sequence before standardization. The server first obtains a localization adjustment parameter. This parameter has a value not less than 1 and is used to adjust the first element sequence. The selection of the localization adjustment parameter is determined based on the actual situation of model training and performance requirements; it can help the model better adapt to different datasets and scenarios. The server uses the obtained localization adjustment parameter to perform optimization on the first element sequence. The specific optimization operation can be multiplication, addition, exponential operation, etc., depending on the nature of the localization adjustment parameter and the model's requirements. The purpose of the optimization operation is to adjust the values ​​of each element in the first element sequence to better match the actual probability distribution of the loose position, thereby improving the model's localization accuracy. For example, if the localization adjustment parameter is a value greater than 1, the server can multiply this value by each element in the first element sequence to increase the values ​​of those elements with larger original values, while relatively decreasing the values ​​of smaller elements. This adjustment allows the model to focus more on the more likely loosening locations, improving the accuracy of identifying these locations. After tuning, the server obtains the third element sequence. Next, the server performs a normalization operation on the third element sequence using a localization activation function. The purpose of normalization is to map the element values ​​in the third element sequence to a specific range (usually between 0 and 1) to more intuitively represent the model's confidence level for each loosening location. Localization activation functions can be common functions such as the softmax function or the sigmoid function. These functions transform the input element values ​​into a probability distribution form, such that the sum of all elements is 1, and the value of each element is between 0 and 1. Through normalization, the server obtains the second element sequence, which is the final inference result. Through the above steps, the server successfully tuned and normalized the first element sequence using localization adjustment parameters, obtaining a more accurate and stable second element sequence. This result will provide valuable information and feedback for subsequent training and optimization.

[0040] In this embodiment of the invention, the tower bolt loosening location model includes a first intermediate network structure and a second intermediate network structure; the aforementioned step of loading the second abnormal soundprint instance into the tower bolt loosening location model, and inferring the sample tower bolt loosening location in the second abnormal soundprint instance through the tower bolt loosening location model to obtain a first inference result can be implemented through the following example.

[0041] (1) Load the second abnormal voiceprint instance into the first intermediate network structure, and perform feature extraction operation on the second abnormal voiceprint instance through the first intermediate network structure to obtain the second feature vector; (2) Load the second feature vector into the second intermediate network structure, and perform scalar multiplication on the second feature vector using the category bias parameter through the second intermediate network structure to obtain the fourth element sequence; wherein, the category bias parameter includes multiple category bias vectors, each category bias vector corresponds to a preset loosening position, and the number of elements in the fourth element sequence is consistent with the number of preset loosening positions; (3) The fourth element sequence is normalized by the localization activation function to obtain the fifth element sequence, and the fifth element sequence is determined as the first inference result; wherein, the elements in the fifth element sequence represent the first confidence value of the preset loose position of the sample iron tower bolt inferred by the tower bolt loose position localization model.

[0042] In this embodiment of the invention, exemplarily, the server has established a model for locating the loose bolt position of a steel tower and is ready to process a second abnormal acoustic signature instance. This model has the same structure as the previously described model, including a first intermediate network structure and a second intermediate network structure. Now, the server will demonstrate in detail how to infer the second abnormal acoustic signature instance through these two network structures and finally obtain a first inference result. The server first loads the second abnormal acoustic signature instance into the first intermediate network structure. This network structure is responsible for extracting features related to bolt loosening from the second abnormal acoustic signature instance. Through a series of convolution, pooling, and fully connected operations, the first intermediate network structure transforms the original acoustic signature data into a second feature vector. This feature vector contains key information extracted from the acoustic signature data, such as frequency changes and amplitude fluctuations, which are crucial for subsequent loosening location inference. Next, the server loads the second feature vector into the second intermediate network structure. This network structure is responsible for inferring the possible location of bolt loosening based on the extracted features. To achieve this goal, the second intermediate network structure uses a set of class bias parameters to perform scalar multiplication on the second feature vector. Each of these class bias parameters corresponds to a preset loosening location. The result of the scalar multiplication operation is a fourth-element sequence, with the number of elements matching the preset number of loosening locations. Each element represents the model's assessment of the probability of loosening at a specific location. This step helps the model map feature vectors to a specific space of loosening locations. To transform the fourth-element sequence into a more interpretable and comparable form, the server uses a localization activation function to normalize it. This activation function maps each element in the sequence to a specific range (e.g., between 0 and 1) to more intuitively represent the model's confidence level for each loosening location. After normalization, the server obtains a fifth-element sequence, the first inference result. This first inference result represents the model's prediction of the loosening location of the sample tower bolts in the second abnormal soundprint instance. The value of each element represents the model's confidence level or confidence value for the loosening at the preset location. These confidence values ​​can be used to evaluate the model's prediction accuracy and provide valuable feedback for subsequent training and optimization. Through the above steps, the server successfully inferred the second abnormal soundprint instance using the tower bolt loosening location localization model and obtained the first inference result regarding the bolt loosening location. This result will provide important information for subsequent analysis and processing.

[0043] In this embodiment of the invention, the aforementioned step of obtaining the learned second cost parameter can be implemented through the following example.

[0044] (1) Obtain the third similarity score between the second feature vector and each of the category deviation vectors; (2) Based on the location confidence value threshold, obtain the similarity score threshold; (3) Determine the deviation between the third similarity score with the maximum value and the similarity score threshold to obtain the third deviation value; (4) Based on the third deviation value, the learned second cost parameter is obtained; wherein the third deviation value and the second cost parameter are proportional.

[0045] In this embodiment of the invention, for example, during the server's execution of locating the loose bolt position of the iron tower, in order to evaluate and optimize the model's performance, it is necessary to calculate the learned second cost parameter. This parameter is obtained based on the similarity score between the second feature vector and each category deviation vector, as well as the location confidence threshold. The server first calculates the similarity between the second feature vector and each category deviation vector, obtaining a series of third similarity scores. These scores reflect the degree of similarity between the second feature vector and the category deviation vector corresponding to each preset loose position. The similarity score can be calculated using methods such as cosine similarity and Euclidean distance, depending on the properties of the feature vector and the category deviation vector, as well as the model's requirements. The server determines a similarity score threshold based on the location confidence threshold. The location confidence threshold is a pre-set standard used to judge whether the model's inference of the loose position is reliable. The similarity score threshold is a score standard corresponding to this threshold, used to evaluate the deviation between the third similarity score and the ideal state. The server finds the score with the maximum value from all third similarity scores and calculates its deviation from the similarity score threshold to obtain the third deviation value. This deviation value reflects the maximum deviation of the model from the ideal state during the inference process. A large third bias value indicates significant errors in the model's inference at certain loose locations; conversely, a small third bias value suggests more accurate inferences. Finally, the server calculates the learned second cost parameter based on the third bias value. This parameter is a quantitative assessment of the model's inference performance, guiding subsequent training and optimization. According to the scenario description, the third bias value is directly proportional to the second cost parameter; that is, the larger the bias value, the larger the cost parameter. This means that when the model experiences large errors during inference, the second cost parameter will increase accordingly, alerting the server to further adjustments and optimizations. Through these steps, the server successfully calculates the learned second cost parameter using the second feature vector, the class bias vector, and the localization confidence threshold. This parameter will provide crucial feedback and guidance for subsequent model training and optimization.

[0046] In this embodiment of the invention, the step of obtaining the learned second cost parameter based on the first confidence value and the location confidence value threshold can be implemented through the following example.

[0047] (1) Determine the deviation between the first confidence value corresponding to each preset loosening position and the positioning confidence value threshold to obtain the second deviation value corresponding to each preset loosening position; (2) If there is a positive second deviation value, the second cost parameter is obtained based on the second deviation value with the maximum value; or, if all the second deviation values ​​are non-positive, the second cost parameter is obtained as 0.

[0048] In this embodiment of the invention, for example, when the server performs the task of locating the loose bolt position of a tower, it has already obtained a first confidence value corresponding to each preset loose position through model inference. To evaluate the accuracy of the model's inference and further optimize model performance, the server needs to calculate a learned second cost parameter. This parameter is calculated based on the deviation between the first confidence value and a positioning confidence value threshold. The server first determines the deviation between the first confidence value and the positioning confidence value threshold for each preset loose position. This deviation reflects the difference between the model's inference result for each loose position and the ideal state. If the first confidence value is higher than the positioning confidence value threshold, it indicates that the model's inference for this position is relatively accurate; if it is lower than the threshold, it indicates that the model's inference for this position may have an error. By calculating the deviation, the server obtains a second deviation value corresponding to each preset loose position. The server then determines the sign of the second deviation value. If there is a positive second deviation value, it indicates that the inference result for at least one loose position is lower than the ideal state, i.e., the model's inference is insufficient. In this case, the server needs to further process these positive deviations to obtain the second cost parameter. If all second deviation values ​​are non-positive (including zero and negative numbers), it indicates that the model's inference results have reached or exceeded the ideal state. In this case, the server can directly set the second cost parameter to 0. If there are positive second deviation values, the server will find the maximum value among these positive values. This maximum positive deviation represents the largest difference between the model's inference of the least accurate position among all preset loose positions and the ideal state. Based on this maximum positive deviation, the server obtains the second cost parameter through certain calculation rules (such as linear mapping, squaring, etc.). This parameter reflects the maximum deviation of the model from the ideal state during the inference process and is an important indicator for evaluating model performance. If all second deviation values ​​are non-positive, it indicates that the model's inference results are generally accurate, with no cases falling below the ideal state. In this case, the server sets the second cost parameter to 0, indicating that the model's inference performance in the current state has reached the expected standard, and no additional optimization is required. Through the above steps, the server calculates the learned second cost parameter based on the first confidence value and the localization confidence value threshold. This parameter not only provides a quantitative indicator for model performance evaluation but also points the way for subsequent model optimization.

[0049] In this embodiment of the invention, the aforementioned step of optimizing and adjusting the model parameters of the tower bolt loosening location model based on the first cost parameter and the second cost parameter to obtain a trained tower bolt loosening location model can be implemented through the following example.

[0050] (1) Detect the number of instances of the first abnormal voiceprint instance of the first abnormal voiceprint instance and the number of instances of the second abnormal voiceprint instance of the second abnormal voiceprint instance in the abnormal voiceprint instance set. (2) Based on the number of instances of the first abnormal voiceprint instance, obtain the first importance weight corresponding to the first cost parameter, and based on the number of instances of the second abnormal voiceprint instance, obtain the second importance weight corresponding to the second cost parameter. (3) Based on the first importance weight and the second importance weight, perform a weighted average calculation on the first cost parameter and the second cost parameter to obtain the final cost parameter; (4) Based on the final cost parameter, optimize and adjust the model parameters of the tower bolt loosening location model to obtain the tower bolt loosening location model after training.

[0051] In this embodiment of the invention, for example, when the server performs the task of locating the loose bolt position of the iron tower, a first cost parameter and a second cost parameter have already been calculated. These two parameters reflect the model's processing performance for different types of abnormal soundprint instances during training. To further optimize the model and improve its accuracy in locating the loose bolt position, the server needs to adjust the model's parameters. This adjustment process is achieved based on the weighted average calculation of the cost parameters. The server first detects the number of first and second abnormal soundprint instances in the abnormal soundprint instance set. This quantity information reflects the distribution of different types of abnormal soundprints in the training set and is the basis for subsequent calculation of importance weights. The server calculates the first importance weight corresponding to the first cost parameter based on the number of instances of the first abnormal soundprint, and calculates the second importance weight corresponding to the second cost parameter based on the number of instances of the second abnormal soundprint. These weights reflect the importance of different types of abnormal soundprints to model training. Generally, the weights corresponding to a larger number of abnormal soundprint types will be relatively higher because they have a greater impact on model performance. The server uses the calculated first and second importance weights to perform a weighted average calculation on the first and second cost parameters. This calculation process comprehensively considers the quantity distribution of different types of anomalous sound signatures and their impact on model performance, thus obtaining a more comprehensive cost evaluation index—the final cost parameter. Based on the calculated final cost parameter, the server optimizes and adjusts the model parameters of the tower bolt loosening location model. This adjustment process may include updating network weights, adjusting the learning rate, etc., aiming to enable the model to more accurately locate the bolt loosening position when processing different types of anomalous sound signatures. After multiple rounds of iteration and optimization, the server obtains a fully trained tower bolt loosening location model, which has significantly improved performance and can more accurately identify and process various anomalous sound signature instances. Through the above steps, the server successfully optimizes and adjusts the tower bolt loosening location model using cost parameters and importance weights. This process not only improves the model's localization accuracy but also lays a solid foundation for subsequent applications and deployments.

[0052] In this embodiment of the invention, the aforementioned step S204 can be implemented through the following examples.

[0053] (1) Obtain the feedback acoustic data of the target tower bolt during the overall monitoring cycle, wherein the overall monitoring cycle consists of multiple monitoring periods, and the monitoring periods include multiple acoustic data collection time nodes; (2) Obtain the initial trend coefficient of the feedback voiceprint data during the monitoring period and the initial baseline value of the feedback voiceprint data during the monitoring period; (3) Based on the feedback voiceprint data at each of the voiceprint acquisition time nodes, the initial baseline value of the feedback voiceprint data, and the initial trend coefficient of the feedback voiceprint data, determine a portion of the fitting error; (4) Update the initial baseline value of the feedback voiceprint data and the initial trend coefficient of the feedback voiceprint data with the optimization goal of reducing the partial fitting error, so as to obtain the first baseline value of the feedback voiceprint data and the first trend coefficient of the feedback voiceprint data. (5) The first trend coefficient of the feedback voiceprint data, the first baseline value of the feedback voiceprint data, and the first difference of the feedback voiceprint data constitute the first voiceprint feature parameter of the feedback voiceprint data. (6) Based on the first voiceprint feature parameters, perform data conversion on the target voiceprint acquisition time node of the monitoring period to obtain partial voiceprint trend data of the monitoring period; (7) Perform an integration operation on partial voiceprint trend data of multiple monitoring periods to obtain overall voiceprint trend data, and perform overall trend analysis on the overall voiceprint trend data to obtain overall voiceprint trend data of the target iron tower bolt; (8) The overall acoustic signature trend data of the target tower bolt is converted into confidence data to obtain the structural damage risk coefficient of the tower structure where the target tower bolt is located.

[0054] In this embodiment of the invention, for example, the server first acquires the feedback acoustic signature data of the target tower bolts within the overall monitoring period. This overall monitoring period consists of multiple consecutive monitoring periods, each containing multiple acoustic signature collection time nodes. The server retrieves all acoustic signature data records of the target tower bolts from the database within the past month (overall monitoring period). This data is grouped by day (monitoring period), and each group contains multiple acoustic signature samples collected at specific time points (acoustic signature collection time nodes). For each monitoring period, the server needs to acquire the initial trend coefficient and initial baseline value of its feedback acoustic signature data. These parameters form the basis for subsequent data analysis. The server performs a preliminary analysis of the acoustic signature data for each monitoring period (e.g., daily), calculating the initial trend coefficient (representing the trend of acoustic signature data changes over time) and the initial baseline value (representing the average level of acoustic signature data) for that period. The server uses the feedback acoustic signature data, initial baseline value, and initial trend coefficient for each acoustic signature collection time node to determine a partial fitting error. This error reflects the deviation between the initial parameters and the actual data. The server calculates partial fitting error by comparing the actual voiceprint data at each time point with the expected value calculated from the initial baseline value and initial trend coefficient. With the goal of reducing partial fitting error, the server updates the initial baseline value and initial trend coefficient of the feedback voiceprint data to obtain more accurate first baseline value and first trend coefficient. Using an optimization algorithm (such as least squares), the server adjusts the initial baseline value and initial trend coefficient to minimize the fitting error between the calculated expected voiceprint data and the actual data. The optimized parameters are called the first baseline value and the first trend coefficient. The server combines the updated first trend coefficient, the first baseline value, and the feedback voiceprint data with the first difference between these two to form the first voiceprint feature parameters. The server integrates the first trend coefficient, the first baseline value, and their differences with the actual voiceprint data (first difference) for each monitoring period to form a first voiceprint feature parameter set containing multiple feature values. Based on the first voiceprint feature parameters, the server performs data transformation on the target voiceprint acquisition time point for each monitoring period to obtain partial voiceprint trend data. This data reflects the changing trend of voiceprints within each monitoring period. The server uses the first voiceprint feature parameters to transform and process the voiceprint data for each monitoring period, extracting key trend information and generating partial voiceprint trend data. This data is represented in charts or numerical form for easy subsequent analysis. The server integrates partial voiceprint trend data from multiple monitoring periods to form overall voiceprint trend data. Through overall trend analysis of this data, the server can understand the voiceprint variation trend of the target tower bolts throughout the entire monitoring period. The server aggregates partial voiceprint trend data from each monitoring period (e.g., daily) to form a continuous overall voiceprint trend dataset.Then, statistical analysis methods are used to perform trend analysis on the overall dataset, identifying the long-term trends and periodic patterns of voiceprint changes. Finally, the server performs confidence data transformation on the overall voiceprint trend data of the target tower bolts to obtain the structural damage risk coefficient of the tower structure. This coefficient reflects the degree of tower structure damage risk based on voiceprint data analysis. The server calculates the structural damage risk coefficient of the tower structure where the target tower bolts are located based on factors such as the fluctuation range, trend, and frequency of outliers in the overall voiceprint trend data. This coefficient is expressed as a percentage or grade and is used to guide subsequent maintenance decisions and safety management measures.

[0055] To more clearly describe the solutions provided in the embodiments of this application, a more detailed explanation is provided below.

[0056] In this embodiment of the invention, for example, the target tower bolts refer to bolts on specific towers that the server has determined may have loosening or other problems. These bolts are the focus of attention and maintenance. Assume a tower network has 100 towers, each with dozens of bolts. The server, through analyzing acoustic signature data, determines that certain bolts on five towers may be loose. These specific bolts on these five towers are referred to as "target tower bolts." The overall monitoring period refers to the total time period during which the server collects and analyzes acoustic signature data on the target tower bolts. This period can be set according to actual needs, such as one month, one quarter, or one year. If the server sets the overall monitoring period to one month, then within this month, the server will periodically collect acoustic signature data about the target tower bolts from sensors installed on the towers. A monitoring period is a smaller time segment within the overall monitoring period. It can be continuous or discontinuous, depending on the data collection and analysis requirements. Acoustic signature data is collected and preliminarily analyzed within each monitoring period. Within a one-month overall monitoring period, the server can set each day as a monitoring period. Thus, there are 30 monitoring periods within a month (assuming the month has 30 days), and acoustic fingerprint data of the target tower bolts is collected and analyzed within each period. The acoustic fingerprint acquisition time node refers to the specific time point during each monitoring period when acoustic fingerprint data is collected. These time points can be fixed (e.g., 3 AM every day) or dynamic (e.g., automatically adjusting the acquisition time based on ambient noise levels). If the server is set to 3 AM every day as the acoustic fingerprint acquisition time node, then at 3 AM during each monitoring period (i.e., every day), the server will collect acoustic fingerprint data about the target tower bolts from sensors installed on the tower. This data will be used for subsequent analysis and processing. By combining the explanations and examples of the above terms, it becomes clearer how the server obtains the feedback acoustic fingerprint data of the target tower bolts throughout the overall monitoring cycle and further processes and analyzes the data to assess the damage risk coefficient of the tower structure.

[0057] Furthermore, the initial trend coefficient refers to a numerical value or parameter indicating the trend of data change, derived by the server after analyzing the collected acoustic fingerprint data of the target tower bolts within a specific monitoring period. This coefficient reflects the overall trend of acoustic fingerprint data over time during that monitoring period, such as rising, falling, or remaining stable. Assume the server collects acoustic fingerprint data of the target tower bolts hourly during a monitoring period (e.g., one day). By analyzing this data, the server can calculate an initial trend coefficient to represent the overall trend of acoustic fingerprint data change within that day. A positive coefficient indicates an overall upward trend in acoustic fingerprint data; a negative coefficient indicates a downward trend; and a coefficient close to zero indicates minimal data change and a stable trend. The initial baseline value refers to a reference level or benchmark value set by the server for the acoustic fingerprint data of the target tower bolts at the beginning of a specific monitoring period. This baseline value is used to compare with subsequently collected acoustic fingerprint data to determine whether the data has undergone abnormal changes or deviated from normal levels. Continuing with the above monitoring period example, suppose at the beginning of a day, the server sets an initial baseline value as a reference level for the acoustic fingerprint data for that day based on historical data or other methods. The server then compares the hourly collected voiceprint data with this baseline value to determine if the data deviates from normal levels. If the data at a certain point in time deviates significantly from the baseline value, the server may consider that there is an anomaly at that point in time, requiring further monitoring and analysis. By combining the initial trend coefficient and the initial baseline value, the server can perform a more comprehensive and accurate analysis of the voiceprint data within the monitoring period, thereby better assessing the condition of the target tower bolts and potential risks. These analysis results will provide important basis for subsequent maintenance decisions and safety management measures.

[0058] Furthermore, the feedback acoustic fingerprint data for each acoustic fingerprint acquisition time point refers to the acoustic fingerprint data collected from the target tower bolts at that specific acquisition time point. Acoustic fingerprint data can be understood as the sound characteristics produced by the bolts at a specific time point; these characteristics may include information about bolt loosening, damage, or other abnormal conditions. Assuming that acoustic fingerprint data is collected hourly during a monitoring period, each hourly time point is an acoustic fingerprint acquisition time point, and the acoustic fingerprint data collected at that time point is the feedback acoustic fingerprint data for that point. This data may include characteristics such as sound frequency and amplitude. Partial fitting error refers to the error that occurs when fitting acoustic fingerprint data using initial baseline values ​​and initial trend coefficients. This error reflects the degree of difference between the model (composed of baseline values ​​and trend coefficients) and the actual data. Suppose there is a simple model that predicts acoustic fingerprint data for each time point based on initial baseline values ​​and initial trend coefficients. However, when this model is used to fit the actually collected data, it is found that the predicted values ​​for some time points differ significantly from the actual values. These differences are the partial fitting error. If the error is large, it indicates that the model may not be accurate enough, and the baseline value or trend coefficient needs to be adjusted to improve the model. By combining the explanations and examples of the above terms, it is possible to better understand how to assess the condition and risk of target tower bolts by analyzing acoustic signature data.

[0059] Furthermore, in data analysis, fitting error refers to the difference between model predictions and actual observations. Saying "reduce some fitting error" means optimizing model parameters (in this case, the initial baseline value and initial trend coefficient) to make the model's predictions closer to the actual observed voiceprint data. Suppose a model predicts the voiceprint data of a tower bolt over a day based on an initial baseline value and an initial trend coefficient. However, when comparing these predictions with the actual collected data, errors are found. To improve the model's accuracy, the baseline value and trend coefficient are adjusted until the error between the predicted and actual values ​​is minimized. Updating the initial baseline value means adjusting it after analyzing the voiceprint data and finding a deviation between the initial baseline value and the actual data, so that it more accurately reflects the characteristics of the actual data. If the initial baseline value is set too high or too low, causing inaccurate model predictions, this baseline value needs to be adjusted based on the actual data. For example, if the initial baseline value is set to 10, but the actual data shows the average is closer to 15, then the baseline value needs to be updated to 15 or a value close to 15. Updating the initial trend coefficient refers to adjusting the initial trend coefficient after analyzing the changing trend of voiceprint data over time to more accurately describe the data's changing trend. Suppose the initial trend coefficient is set to a positive number, indicating that the voiceprint data shows an upward trend over time. However, actual data analysis shows that the voiceprint data is gradually decreasing. In this case, the trend coefficient needs to be updated to a negative number to more accurately reflect the downward trend of the data. Through an optimization process (i.e., reducing the fitting error), updated baseline values ​​and trend coefficients can be obtained, referred to here as the "first baseline value" and "first trend coefficient." These are more accurate model parameters obtained by iteratively adjusting the initial parameters. After a series of adjustments and optimizations, it may be found that when the baseline value is set to 17 and the trend coefficient is set to -0.2, the model's prediction results best match the actual voiceprint data. These two values ​​(17 and -0.2) are what are referred to as the "first baseline value" and "first trend coefficient." In summary, by reducing the fitting error as the optimization objective, the initial baseline value and initial trend coefficient can be updated to obtain more accurate first baseline values ​​and first trend coefficients for subsequent data analysis and prediction.

[0060] Furthermore, the first trend coefficient, obtained after optimization, is a parameter used to describe the overall trend of acoustic signature data over time. This parameter is determined by analyzing the acoustic signature data and reducing fitting errors. It reflects the main trend of data change during the monitoring period, such as rising, falling, or remaining stable. Assuming the optimized first trend coefficient is -0.5, this means that the acoustic signature data of the target tower bolts generally shows a downward trend throughout the monitoring period. This coefficient can help predict future data changes and provide a reference for subsequent maintenance decisions. The first baseline value, determined during optimization, represents the reference level or benchmark value of the acoustic signature data at a specific point in time. This value is obtained by analyzing actual acoustic signature data and adjusting the initial baseline value. It reflects the average level or normal state of the data during the monitoring period. Assuming the optimized first baseline value is 20, this means that the acoustic signature data of the target tower bolts should be close to this value under average or normal conditions throughout the monitoring period. If the actual data deviates too far from this baseline value, it may indicate an anomaly requiring further attention and analysis. The first difference refers to the difference between the actual voiceprint data at each voiceprint collection time point and the data predicted based on the first baseline value and the first trend coefficient. This difference reflects the degree of deviation between the model prediction and the actual data and is an important indicator for evaluating the accuracy of the model. Suppose that at a certain voiceprint collection time point, the actual voiceprint data collected is 25, while the data predicted based on the first baseline value and the first trend coefficient is 23. Then, the first difference at this time point is 2 (i.e., 25-23). ​​If the difference is large, it may indicate that the model prediction is inaccurate or that there are other unknown factors affecting the data. The first voiceprint feature parameter is composed of three parameters: the first trend coefficient, the first baseline value, and the first difference, which are used to comprehensively describe the voiceprint data characteristics of the target tower bolts during a specific monitoring period. This parameter set can help to understand the inherent laws and potential risks of the data more deeply. By combining the first trend coefficient (-0.5), the first baseline value (20), and the first difference at each time point (such as 2 in the example above), a complete first voiceprint feature parameter set can be obtained. This parameter set can be used for subsequent data analysis, anomaly detection, or risk assessment tasks. For example, if the first difference at a certain time point is much greater than that at other time points, then this time point may be considered a potential outlier that requires further attention and analysis.

[0061] Furthermore, the target voiceprint acquisition time node during the monitoring period refers to the time point within a specific monitoring period where voiceprint data was collected. These time nodes are usually pre-set for collecting and analyzing voiceprint data. Assuming a one-hour monitoring period is set, and voiceprint data is collected every minute, then each minute within this monitoring period is a target voiceprint acquisition time node. Data transformation refers to the process of converting raw data or feature parameters into a new data form or expression using a specific algorithm or model. In voiceprint data analysis, data transformation may involve processing, transforming, or encoding the raw voiceprint data. In this example, data transformation may refer to processing the raw voiceprint data using a first voiceprint feature parameter (including a first trend coefficient, a first baseline value, and a first difference) to extract more meaningful voiceprint trend data. This process may involve steps such as data smoothing, denoising, and feature extraction. Partial voiceprint trend data refers to a portion of the data obtained during the data transformation process that describes the trend of voiceprint data over time. This data is usually extracted from the raw voiceprint data and used for further analysis and prediction. By transforming the raw voiceprint data from the target voiceprint acquisition time points during the monitoring period, a series of data points describing the voiceprint variation trend can be obtained. These data points constitute partial voiceprint trend data, which can help to more clearly understand the variation patterns and potential trends of voiceprint data. For example, if partial voiceprint trend data shows that the voiceprint amplitude is gradually increasing, it may mean that there is a risk of loosening or damage to the bolts of the target tower. By using the first voiceprint feature parameters to transform the data from the target voiceprint acquisition time points during the monitoring period, partial voiceprint trend data can be obtained, which provides an important basis for subsequent voiceprint analysis and prediction.

[0062] Furthermore, in voiceprint monitoring, to gain a more comprehensive understanding of the target tower bolts' condition, data is typically collected over multiple different time periods. These time periods are called monitoring periods. Each monitoring period contains a certain number of voiceprint collection time points and corresponding voiceprint data. Assuming a week-long voiceprint monitoring operation is conducted on the target tower bolts, with a fixed time period (e.g., 9 AM to 10 AM) selected each day for data collection, then this time period each day of the week can be considered a monitoring period, resulting in a total of seven monitoring periods. Integration refers to the process of merging, aligning, and unifying data from multiple sources or different time periods. In voiceprint monitoring, integration typically involves merging partial voiceprint trend data from multiple monitoring periods into a continuous dataset for more comprehensive analysis. Integrating the partial voiceprint trend data from the seven monitoring periods mentioned earlier means merging the data from these seven periods into a complete dataset containing the voiceprint change trend throughout the entire monitoring period (one week). This process may involve data interpolation, smoothing, and other operations to ensure data continuity and consistency. The overall voiceprint trend data refers to the dataset obtained after integration, reflecting the voiceprint change trend throughout the entire monitoring period. These datasets contain information from multiple monitoring periods, providing a more comprehensive and macroscopic perspective for observing and analyzing the condition of the target tower bolts. By integrating partial acoustic signature trend data from seven monitoring periods, a comprehensive acoustic signature trend dataset containing daily acoustic signature variation trends over a week can be obtained. This dataset can help identify long-term acoustic signature variation patterns, periodic fluctuations, or abnormal events. Overall trend analysis refers to the process of using statistical, signal processing, or machine learning methods to conduct in-depth analysis of overall acoustic signature trend data to reveal long-term trends, periodic changes, and abnormal events within the acoustic signature data. This analysis helps assess the overall condition of the target tower bolts and predict future trends. Overall trend analysis of the overall acoustic signature trend data can employ methods such as time series analysis, spectral analysis, or machine learning algorithms to extract useful information from the data. For example, time series analysis can identify seasonal variations, trend terms, and random fluctuations in the acoustic signature data; machine learning algorithms can classify or predict acoustic signature data to discover potential failure modes or provide early warnings of abnormal events.

[0063] Furthermore, assuming continuous monitoring for a week yields acoustic signature data of the target tower bolts, and after processing and analysis, a dataset describing its overall trend is obtained. This dataset may include changes in acoustic signature amplitude, frequency, and other characteristics over time. Confidence data transformation refers to converting the raw data (here, the overall acoustic signature trend data) into a format representing the data's reliability or confidence level. Confidence is a crucial consideration in risk assessment and prediction, reflecting the credibility of the assessment results. For the overall acoustic signature trend data of the target tower bolts, statistical methods, machine learning models, or other algorithms can be used to calculate the confidence level of each data point. For example, if the acoustic signature data at a certain point in time deviates significantly from the normal pattern, but the model's prediction confidence for that data point is low, then a more cautious decision-making process may be adopted, as this implies that the model's prediction for that data point is less certain. The structural damage risk coefficient is a quantitative indicator used to assess the risk of damage to the tower structure where the target tower bolts are located. This coefficient is typically calculated based on multiple factors, including acoustic signature data, environmental factors, and historical data. Assuming that confidence levels are obtained for the acoustic signature data at each time point after confidence data transformation, this confidence data, combined with other relevant information (such as environmental factors and historical damage records), can be used to calculate the structural damage risk coefficient of the tower structure where the target tower bolt is located using a risk assessment model. This coefficient may be a value between 0 and 1, with a higher value indicating a higher risk of damage. In summary, by transforming the overall acoustic signature trend data of the target tower bolts into confidence data and combining it with other relevant information, the structural damage risk coefficient of the tower structure can be calculated, thus providing support for subsequent maintenance and management decisions.

[0064] In this embodiment of the invention, the aforementioned step of obtaining the feedback acoustic data of the target tower bolt during the overall monitoring cycle can be implemented through the following example.

[0065] (1) Obtain the original feedback acoustic data of the target tower bolts during the monitoring period; (2) Perform data denoising operation on the original feedback acoustic data of the target tower bolt during the monitoring period to obtain the denoised feedback acoustic data during the monitoring period; (3) Perform a data dimensionality reduction operation on the denoised feedback voiceprint data to obtain the dimensionality-reduced feedback voiceprint data for the monitoring period; (4) Perform data normalization operation on the dimension-reduced feedback voiceprint data of the monitoring period to obtain the feedback voiceprint data of the monitoring period; (5) The feedback voiceprint data of multiple monitoring periods are used to form the feedback voiceprint data of the overall monitoring cycle.

[0066] In this embodiment of the invention, for example, the server first establishes a communication connection with the acoustic signature acquisition devices deployed on the target tower, receiving the raw acoustic signature data transmitted by these devices in real time or at regular intervals. This data is unprocessed and directly reflects the sound vibration of the target tower bolts during a specific monitoring period. The server stores this raw data in a dedicated data warehouse for subsequent processing and analysis. Since the raw acoustic signature data may contain irrelevant information such as environmental noise and equipment interference, the server needs to process this data using data denoising algorithms. For example, the server can use spectral analysis to identify and filter out frequency components that are clearly not part of the acoustic signature characteristics of the target tower bolts. After data denoising, the server obtains cleaner, denoised feedback acoustic signature data that primarily reflects the acoustic signature characteristics of the target tower bolts. When processing large-scale acoustic signature data, in order to improve computational efficiency and reduce storage costs, the server needs to perform dimensionality reduction processing on the denoised data. For example, the server can use dimensionality reduction algorithms such as Principal Component Analysis (PCA) to extract the main feature components in the raw data and map the high-dimensional data to a low-dimensional space. In this way, the server can greatly reduce the complexity of the data and the amount of computation while retaining the main information. To eliminate data inconsistencies caused by differences in acquisition conditions and equipment between different monitoring periods, the server needs to normalize the dimensionality-reduced data. For example, the server can use a min-max normalization method to scale the value of each feature to a uniform range (e.g., between 0 and 1). This makes the data from different monitoring periods comparable, facilitating subsequent overall analysis and trend prediction. Finally, the server needs to integrate the processed feedback acoustic fingerprint data from different monitoring periods to form a complete dataset covering the entire monitoring cycle. This dataset is the foundation for subsequent overall trend analysis, structural damage risk assessment, and other tasks. The server can use time series analysis to connect these time-series data points to form a continuous data curve or model reflecting the acoustic fingerprint change trend of the target tower bolts throughout the entire monitoring cycle.

[0067] In this embodiment of the invention, the aforementioned step of obtaining the original feedback acoustic data of the target tower bolt during the monitoring period can be implemented through the following example.

[0068] (1) Obtain the frequency characteristics, amplitude characteristics and time domain characteristics of the target tower bolt at the time node of acoustic text acquisition; (2) The frequency characteristics, amplitude characteristics and time domain characteristics of the target tower bolt at the soundprint acquisition time node constitute the original feedback soundprint data of the target tower bolt at the soundprint acquisition time node; (3) The original feedback acoustic data of the target tower bolt at multiple acoustic data acquisition time nodes constitute the original feedback acoustic data of the target tower bolt during the monitoring period.

[0069] In this embodiment of the invention, for example, the server receives data from acoustic signature acquisition devices deployed on the target tower. At each acoustic signature acquisition time point, these devices capture the sound signal emitted by the bolt and extract the frequency characteristics of the sound using techniques such as spectrum analysis. Frequency characteristics reflect the distribution and intensity of different frequency components in the sound signal and are an important component of the acoustic signature data. The server records this frequency characteristic data as part of the raw feedback acoustic signature data. In addition to frequency characteristics, the server also needs to acquire the amplitude characteristics of the sound signal. Amplitude characteristics describe the magnitude of the sound signal fluctuation, i.e., the loudness of the sound. The acoustic signature acquisition devices measure and record the amplitude values ​​of the sound signal at each time point; these values ​​reflect the vibration intensity of the bolt at different time points. The server also includes this amplitude characteristic data in the raw feedback acoustic signature data. Temporal characteristics refer to the changing features of the sound signal along the time axis, including the signal duration and waveform. The server extracts the temporal characteristics by analyzing the sound signals captured by the acoustic signature acquisition devices at each time point. These characteristics help to understand the evolution of the bolt acoustic signature over time, which is of great significance for subsequent fault detection and prediction. The server integrates the frequency, amplitude, and temporal characteristics data acquired at each acoustic signature acquisition time point to form a complete data point. These data points collectively constitute the raw feedback acoustic signature data of the target tower bolt at that time point. This data not only contains the frequency and loudness information of the bolt's sound but also reflects its temporal variation characteristics. The server aggregates the raw feedback acoustic signature data from all acoustic signature acquisition time points within a monitoring period to form a continuous dataset. This dataset records the acoustic signature changes of the bolt throughout the monitoring period, providing a foundation for subsequent data processing and analysis. Through in-depth analysis of this data, the server can assess the bolt's working condition, predict potential failure risks, and provide targeted maintenance recommendations to maintenance personnel.

[0070] In this embodiment of the invention, the original feedback acoustic data of the target tower bolt during the monitoring period includes the original feedback acoustic data of the target tower bolt at each acoustic data acquisition time node during the monitoring period; the aforementioned step of performing data denoising operation on the original feedback acoustic data of the target tower bolt during the monitoring period to obtain the denoised feedback acoustic data of the monitoring period can be implemented through the following example.

[0071] (1) The original feedback acoustic data of the target iron tower bolt in each acoustic data acquisition time node during the monitoring period that are missing acoustic data and acoustic noise data are screened out to obtain the original feedback acoustic data after noise reduction. (2) Perform redundant data elimination operation on the original feedback voiceprint data of each voiceprint acquisition time node in the denoised original feedback voiceprint data to obtain the denoised feedback voiceprint data of the monitoring period.

[0072] In this embodiment of the invention, for example, after the server receives the raw feedback acoustic data of the target tower bolt at each acoustic data acquisition time point during the monitoring period, the first step is data cleaning. The server iterates through these data, checking for missing or outlier values. Missing values ​​may be due to data loss caused by acquisition equipment failure, communication interruption, etc.; while outliers may be noise data caused by non-target sound sources such as environmental noise or equipment interference. The server uses preset rules or algorithms, such as threshold-based judgment and statistical analysis, to identify and filter out these missing acoustic data and acoustic noise data. This ensures the accuracy and reliability of subsequent analysis and avoids misjudgments or misleading information due to data quality issues. After filtering out missing and noise data, the server needs to further process the remaining raw feedback acoustic data to eliminate potentially redundant data. Redundant data refers to data collected at multiple acoustic data acquisition time points that are highly similar or repetitive. This data has no additional value for subsequent analysis and may even increase computational burden and storage costs. The server can use algorithms such as data clustering and similarity comparison to identify and eliminate this redundant data. For example, the server can extract features from the voiceprint data at each time point and then calculate the similarity or distance between data from different time points. Based on a preset similarity threshold or distance standard, the server can determine which data is redundant and retain only one or a few of the most representative data points. After these two steps, the server obtains denoised, cleaner, and more concise feedback voiceprint data of the target tower bolts during the monitoring period. This data will serve as the basis for subsequent analysis and prediction. It should be noted that in practical applications, the server may flexibly adjust and optimize the above steps according to specific business needs and data analysis objectives. For example, the server can select appropriate denoising algorithms and parameter settings based on different monitoring scenarios and data characteristics; it can also combine other data sources or prior knowledge to assist in the data cleaning and denoising process.

[0073] In this embodiment of the invention, the dimensionality-reduced feedback voiceprint data for the monitoring period includes dimensionality-reduced feedback voiceprint data for each voiceprint acquisition time node in the monitoring period across multiple voiceprint feature dimensions; the aforementioned step of performing data normalization operations on the dimensionality-reduced feedback voiceprint data for the monitoring period to obtain the feedback voiceprint data for the monitoring period can be implemented through the following example.

[0074] (1) Obtain the maximum and minimum dimensionality-reduced feedback voiceprint data among the dimensionality-reduced feedback voiceprint data of the voiceprint feature dimension at multiple voiceprint acquisition time nodes; (2) Obtain the first difference value between the maximum dimension-reduced feedback voiceprint data and the minimum dimension-reduced feedback voiceprint data; (3) Obtain the second difference value between the dimensionality-reduced feedback voiceprint data of the voiceprint feature dimension at each of the voiceprint acquisition time nodes and the minimum dimensionality-reduced feedback voiceprint data; (4) The ratio between the second difference value and the first difference value corresponding to each of the voiceprint acquisition time nodes is used as the standardized feedback voiceprint data of the voiceprint acquisition time node in the voiceprint feature dimension; (5) The standardized feedback voiceprint data of the voiceprint acquisition time node at multiple voiceprint feature dimensions are used to form the feedback voiceprint data of the voiceprint acquisition time node; (6) The feedback voiceprint data of multiple voiceprint acquisition time nodes are used to form the feedback voiceprint data of the monitoring period.

[0075] In this embodiment of the invention, for example, in the server, for the dimensionality-reduced feedback voiceprint data of each monitoring period, the analysis is first performed on each voiceprint feature dimension. The server iterates through the dimensionality-reduced data of all voiceprint acquisition time nodes in that feature dimension within the monitoring period, finding the maximum and minimum values. These values ​​represent the range and fluctuation of the data in that feature dimension. The server then calculates the difference between the maximum and minimum dimensionality-reduced feedback voiceprint data in each voiceprint feature dimension, i.e., the first difference value. This difference value reflects the dispersion or range of change of the data in that feature dimension. For each voiceprint acquisition time node, the server calculates the difference between its dimensionality-reduced data in each voiceprint feature dimension and the minimum dimensionality-reduced data in that feature dimension, i.e., the second difference value. This difference value represents the offset or degree of change of the data at that time node relative to the minimum value in that feature dimension. The server calculates the ratio of the second difference value of each voiceprint acquisition time node to the first difference value in that feature dimension, and the result is the standardized feedback voiceprint data of that time node in that feature dimension. This process essentially normalizes the data, mapping it to a uniform range (usually between 0 and 1) to eliminate the influence of differences in units and orders of magnitude between different feature dimensions. For each acoustic signature acquisition time point, the server integrates its standardized feedback acoustic signature data across multiple acoustic signature feature dimensions, forming a multi-dimensional data point. This data point comprehensively reflects the performance of bolt acoustic signatures across different feature dimensions at that time point. Finally, the server aggregates the feedback acoustic signature data from all acoustic signature acquisition time points within a monitoring period, forming a complete dataset. This dataset records the acoustic signature changes of bolts across different feature dimensions throughout the entire monitoring period, providing a crucial data foundation for subsequent structural damage risk assessment and prediction.

[0076] In this embodiment of the invention, the aforementioned step of determining a portion of the fitting error based on the feedback voiceprint data at each of the voiceprint acquisition time nodes, the initial baseline value of the feedback voiceprint data, and the initial trend coefficient of the feedback voiceprint data can be implemented through the following example.

[0077] (1) Based on the initial baseline value of the feedback voiceprint data and the initial trend coefficient of the feedback voiceprint data, perform one-dimensional data transformation on the voiceprint acquisition time node to obtain the first transformation output; (2) When the error value between the first conversion output and the feedback voiceprint data at the voiceprint acquisition time node is lower than the tolerance range threshold, obtain the partial fitting error of the second growth relationship between the voiceprint acquisition time node and the error value. (3) When the error value between the first conversion output and the feedback voiceprint data of the voiceprint acquisition time node is greater than the tolerance range threshold, a first benchmark value with the second growth relationship of the tolerance range threshold is obtained, the error value is multiplied by the tolerance range threshold, and the time trend coefficient is subtracted from the first benchmark value to obtain the voiceprint acquisition time node partial fitting error. (4) The partial fitting errors of the multiple voiceprint acquisition time nodes are summed to obtain the partial fitting error.

[0078] In this embodiment of the invention, for example, the server first obtains the initial baseline value and initial trend coefficient of the feedback voiceprint data. The initial baseline value represents the baseline level of the voiceprint data at the start of monitoring, while the initial trend coefficient reflects the overall trend of the voiceprint data over time.

[0079] Next, the server performs a metadata transformation on the feedback voiceprint data at each voiceprint acquisition time point. This transformation process is based on an initial baseline value and an initial trend coefficient, aiming to convert the original voiceprint data into a form that is easier to analyze and compare. The result of the transformation is called the first transformation output. The server calculates the error value between the first transformation output and the original feedback voiceprint data. This error value reflects the degree of difference between the transformed data and the original data. The server pre-sets a tolerance range threshold to determine whether the error value is within an acceptable range. If the error value is below this threshold, it indicates that the transformed data is relatively close to the original data, and the fitting effect is good; if the error value is above this threshold, it indicates that the fitting effect is poor and further processing is required. When the error value is below the tolerance range threshold, the server obtains a partial fitting error of the quadratic growth relationship between the voiceprint acquisition time point and the error value. This means that the error value will be further processed to reflect its trend over time. When the error value is above the tolerance range threshold, the server adopts different processing methods. First, the server obtains a first baseline value of the quadratic growth relationship with the tolerance range threshold. Then, the error value is multiplied by the tolerance range threshold to amplify its impact. Next, the time trend coefficient is subtracted from the first baseline value to obtain the partial fitting error for each voiceprint acquisition time point. This processing method can more accurately reflect the error situation when the data fitting effect is poor. Finally, the server sums the partial fitting errors of multiple voiceprint acquisition time points to obtain the total partial fitting error. This total error value reflects the overall data fitting effect over the entire monitoring period, providing an important basis for subsequent analysis and decision-making.

[0080] In this embodiment of the invention, the target voiceprint acquisition time node includes the first voiceprint acquisition time node and the last voiceprint acquisition time node. The aforementioned step of converting the target voiceprint acquisition time node of the monitoring period based on the first voiceprint feature parameter to obtain partial voiceprint trend data of the monitoring period can be implemented through the following example.

[0081] (1) Perform a multiplication operation between the first trend coefficient in the first voiceprint feature parameter and the first voiceprint acquisition time node in the monitoring period to obtain the first time trend coefficient; (2) Perform an addition operation between the first time trend coefficient and the first baseline value to obtain partial voiceprint trend data corresponding to the first voiceprint acquisition time node; (3) Perform a multiplication operation between the first trend coefficient in the first voiceprint feature parameter and the last voiceprint acquisition time node in the monitoring period to obtain the second time trend coefficient; (4) Perform an addition operation between the second time trend coefficient and the first baseline value to obtain partial voiceprint trend data corresponding to the last voiceprint acquisition time node; (5) The partial voiceprint trend data corresponding to the first voiceprint collection time node and the partial voiceprint trend data corresponding to the last voiceprint collection time node constitute the partial voiceprint trend data of the monitoring period.

[0082] In this embodiment of the invention, exemplarily, the server first obtains first voiceprint feature parameters, including a first trend coefficient and a first baseline value. The first trend coefficient represents the overall trend of voiceprint data over time, while the first baseline value is the baseline level of voiceprint data at the start of monitoring. The server identifies the first voiceprint acquisition time node in the monitoring period. This node is the starting point of the monitoring period, and the corresponding voiceprint data reflects the state of the target tower bolts at the start of monitoring. Next, the server performs a multiplication operation between the first trend coefficient and the first voiceprint acquisition time node to obtain a first time trend coefficient. This coefficient reflects the changing trend of voiceprint data from the start of monitoring to the first time node. Then, the server performs an addition operation between the first time trend coefficient and the first baseline value to obtain partial voiceprint trend data corresponding to the first voiceprint acquisition time node. This data is a conversion result combining the baseline value and the trend coefficient, used to represent the predicted value or trend value of the voiceprint data at that time node. Similarly, the server identifies the last voiceprint acquisition time node in the monitoring period. This node marks the end of the monitoring period, and the corresponding acoustic signature data reflects the state of the target tower bolts at the end of the monitoring. The server multiplies the first trend coefficient with the last acoustic signature acquisition time node to obtain the second time trend coefficient. This coefficient reflects the trend of acoustic signature data changes from the start of monitoring to the last time node. Then, the server adds the second time trend coefficient with the first baseline value to obtain partial acoustic signature trend data corresponding to the last acoustic signature acquisition time node. Similarly, this data is a conversion result combining the baseline value and the trend coefficient, used to represent the predicted or trend value of the acoustic signature data at that time node. Finally, the server integrates the partial acoustic signature trend data corresponding to the first and last acoustic signature acquisition time nodes to form the partial acoustic signature trend data for the monitoring period. These data reflect the changing trends and predicted values ​​of acoustic signature data at the beginning and end of the monitoring period, providing important reference for subsequent structural damage risk assessment and prediction. In this embodiment of the invention, the aforementioned step of performing an overall trend analysis on the overall acoustic signature trend data to obtain the overall acoustic signature trend data of the target tower bolt can be implemented through the following example.

[0083] (1) Perform univariate regression on the overall voiceprint trend data to obtain the second trend coefficient of the feedback voiceprint data and the second baseline value of the feedback voiceprint data; (2) The second trend coefficient of the feedback voiceprint data, the second baseline value of the feedback voiceprint data, and the second error of the feedback voiceprint data constitute the second voiceprint feature parameter of the feedback voiceprint data; (3) Based on the second voiceprint feature parameters, perform data conversion to obtain the overall voiceprint trend data of the overall monitoring period.

[0084] In this embodiment of the invention, for example, the server first collects overall acoustic signature trend data of the target tower bolts over the entire monitoring period. This data reflects the changes in the acoustic signature characteristics of the bolts throughout the monitoring period.

[0085] To analyze the overall trend of this data, the server employs univariate regression. Univariate regression is a statistical technique used to study the relationship between two variables, one of which is the dependent variable (in this case, voiceprint data), and the other is the independent variable (in this case, time). Through univariate regression, the server can fit a straight line or curve to describe the trend of voiceprint data over time. In this process, the server calculates two important parameters: the second trend coefficient and the second baseline value. The second trend coefficient reflects the slope or rate of change of the voiceprint data over time, while the second baseline value represents the baseline level or intercept of the voiceprint data at the start of monitoring. The server combines the calculated second trend coefficient, the second baseline value, and the second error of the feedback voiceprint data to form the second voiceprint feature parameters of the feedback voiceprint data. These parameters comprehensively describe the characteristics of the voiceprint data throughout the overall monitoring period, including its baseline level, trend, and error range. The second error refers to the residual or fitting error generated during the univariate regression process, reflecting the degree of difference between the actual voiceprint data and the regression line. By incorporating the second error into the second voiceprint feature parameters, the server can more accurately assess the quality and reliability of the data fit. Finally, the server utilizes the second voiceprint feature parameters to transform the overall voiceprint trend data. This transformation process aims to convert the raw voiceprint data into a more easily analyzed and compared form, providing a more intuitive representation of the overall trend. Specifically, the server may apply the second trend coefficient and the second baseline value to a mathematical model or formula to calculate the overall voiceprint trend data for the entire monitoring period. This data can be predicted values ​​at each time point, points on the trend line, or other relevant indicators used to describe the changes and trends of the voiceprint data throughout the monitoring period. Through this processing and analysis workflow, the server can accurately obtain the overall voiceprint trend data of the target tower bolts within the entire monitoring period, providing strong data support for subsequent structural health assessments and predictions.

[0086] In this embodiment of the invention, before the aforementioned step of converting the overall acoustic signature trend data of the target tower bolt into confidence data to obtain the structural damage risk coefficient of the target tower bolt, this embodiment of the invention also provides the following implementation method.

[0087] (1) Obtain the feedback acoustic data of the entire monitoring period before the actual tower structure is damaged as the target sample instance, and obtain the feedback acoustic data of the entire monitoring period after the actual tower structure is repaired as the non-target sample instance. (2) Obtain the overall voiceprint trend data corresponding to the target sample instance and the overall voiceprint trend data corresponding to the non-target sample instance; (3) Generate the numerical region of the target sample instance based on the overall voiceprint trend data of the target sample instance, and generate the numerical region of the non-target sample instance based on the overall voiceprint trend data of the non-target sample instance. (4) The overlapping part between the numerical region of the target sample instance and the numerical region of the non-target sample instance is taken as the risk numerical region; The aforementioned step of converting the overall acoustic signature trend data of the target tower bolts into confidence data to obtain the structural damage risk coefficient of the target tower bolts can be implemented through the following example.

[0088] (1) When the overall acoustic trend data of the target tower bolt is in the risk value area, the confidence level corresponding to the risk value area is used as the structural damage risk coefficient of the tower structure where the target tower bolt is located.

[0089] In this embodiment of the invention, exemplarily, in a real-world scenario, the process by which the server performs confidence data conversion on the overall acoustic signature trend data of the target tower bolts to obtain the structural damage risk coefficient can be described in detail as follows: First, the server needs to collect acoustic signature data of the actual tower structure before and after damage as samples. This is typically done by installing acoustic signature monitoring equipment on the tower and collecting acoustic signature feedback data of the structure under different conditions. Specifically, the server acquires the feedback acoustic signature data of the actual tower during the overall monitoring period before structural damage, as the target sample instance; simultaneously, the server also acquires the feedback acoustic signature data of the actual tower during the overall monitoring period after structural maintenance, as the non-target sample instance. Next, the server processes this acoustic signature data and extracts the overall acoustic signature trend data. The overall acoustic signature trend data describes the changes in acoustic signature of the tower structure at different stages, reflecting the health status of the tower structure. The server acquires the overall acoustic signature trend data for the corresponding target sample instance and non-target sample instance respectively. Then, the server generates corresponding numerical regions based on this overall acoustic signature trend data. For target sample instances, their overall acoustic signature trend data constitutes a specific numerical region; for non-target sample instances, their overall acoustic signature trend data also constitutes a different numerical region. These numerical regions are the server's quantitative representation of the acoustic signature characteristics of the tower structure under different states. The server then identifies the overlapping portion between these two numerical regions and designates it as a risk numerical region. The risk numerical region represents the range of states where the tower structure may be at risk of damage. This means that if the overall acoustic signature trend data of the target tower bolt falls within this region, then the tower structure in which it is located faces a certain risk of structural damage. Finally, when the server processes the overall acoustic signature trend data of the target tower bolt, it compares the target tower bolt's data with the risk numerical region. If the overall acoustic signature trend data of the target tower bolt falls within the risk numerical region, the server uses the confidence level of the corresponding risk numerical region as the structural damage risk coefficient of the tower structure in which the target tower bolt is located. This risk coefficient helps engineers assess the health status of the tower structure, promptly identify potential safety hazards, and take appropriate maintenance measures. Throughout the process, the server collects acoustic signature data from real towers, extracts overall acoustic signature trends, generates numerical regions, identifies risk numerical regions, and finally compares the data of the target tower's bolts with the risk numerical regions to obtain the structural damage risk coefficient. This process not only improves the accuracy of monitoring the health status of tower structures but also provides strong support for tower maintenance and safety management.

[0090] This invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned method for locating and maintaining the loose position of iron tower bolts based on deep learning and voiceprint analysis. Figure 2 As shown, Figure 2 This is a structural block diagram of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To enable data transmission or interaction, the memory 111, processor 112, and communication unit 113 are electrically connected to each other directly or indirectly. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

[0091] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the foregoing illustrative discussions are not intended to be exhaustive or to limit the present disclosure to the precise forms disclosed. Numerous modifications and variations are possible in accordance with the foregoing teachings. These embodiments were chosen and described in order to best illustrate the principles of the present disclosure and its practical application, thereby enabling those skilled in the art to best utilize the disclosure and to employ various embodiments with different modifications to suit a particular intended application.

Claims

1. A method for locating and maintaining loose bolts on iron towers based on deep learning and voiceprint analysis, characterized in that... include: Acquire current feedback acoustic data for tower bolts based on preset sensors; The pre-trained tower bolt loosening location model is called to process the current feedback voiceprint data to obtain the current tower bolt loosening location; Identify the target tower bolt corresponding to the current location of the loose tower bolt; The feedback acoustic data of the target tower bolt during the overall monitoring cycle is obtained, and the structural damage risk coefficient of the tower structure where the target tower bolt is located is calculated. Based on the structural damage risk coefficient, a maintenance strategy for the target tower bolts is determined. The tower bolt loosening location model is obtained through the following methods: Obtain an abnormal soundprint instance set; the abnormal soundprint instance set includes multiple sets of first abnormal soundprint instances and multiple sets of second abnormal soundprint instances, wherein each set of first abnormal soundprint instances includes feedback soundprint data of the sample tower bolts being loose at a preset loose position, and each set of second abnormal soundprint instances includes other feedback soundprint data of the sample tower bolts being loose at other positions. The first abnormal voiceprint instance is loaded into the tower bolt loosening location model to perform the training process and obtain the first learned cost parameter. The second abnormal soundprint instance is loaded into the tower bolt loosening location model. The tower bolt loosening location model is used to infer the loosening location of the sample tower bolt in the second abnormal soundprint instance to obtain a first inference result. The first inference result includes the first confidence value of each preset loosening location inferred by the tower bolt loosening location model. Based on the first confidence value and the location confidence value threshold, the learned second cost parameter is obtained; The number of instances of the first abnormal voiceprint instance of the first abnormal voiceprint instance and the number of instances of the second abnormal voiceprint instance of the second abnormal voiceprint instance are detected in the abnormal voiceprint instance set. Based on the number of instances of the first abnormal voiceprint instance, the first importance weight corresponding to the first cost parameter is obtained, and based on the number of instances of the second abnormal voiceprint instance, the second importance weight corresponding to the second cost parameter is obtained. Based on the first importance weight and the second importance weight, a weighted average calculation is performed on the first cost parameter and the second cost parameter to obtain the final cost parameter; Based on the final cost parameter, the model parameters of the tower bolt loosening location model are optimized and adjusted to obtain a trained tower bolt loosening location model. The abnormal voiceprint instance set also includes a target variable corresponding to the first abnormal voiceprint instance. The target variable is used to indicate the ground truth value of the sample tower bolt loosening location in the first abnormal voiceprint instance. The tower bolt loosening location model includes a first intermediate network structure and a second intermediate network structure. The step of loading the first abnormal voiceprint instance into the tower bolt loosening location model and performing the training process to obtain the learned first cost parameter includes: The first abnormal voiceprint instance is loaded into the first intermediate network structure, and a feature extraction operation is performed on the first abnormal voiceprint instance through the first intermediate network structure to obtain the first feature vector. The first feature vector is loaded into the second intermediate network structure, and the first feature vector is multiplied by a scalar multiplication operation using a category bias parameter through the second intermediate network structure to obtain a first element sequence; wherein, the category bias parameter includes multiple category bias vectors, each category bias vector corresponds to a preset loosening position, and the number of elements in the first element sequence is the same as the number of preset loosening positions; Obtain the positioning adjustment parameters; the value of the positioning adjustment parameters shall not be less than 1. The third element sequence is obtained by performing an optimization operation on the first element sequence using the positioning adjustment parameters. The third element sequence is normalized by performing a localization activation function to obtain a second element sequence, which is then determined as the second inference result. The elements in the second element sequence represent the second confidence value of the preset loosening position inferred by the tower bolt loosening position localization model for the sample tower bolt. The second inference result includes the second confidence value of each preset loosening position inferred by the tower bolt loosening position localization model for the sample tower bolt becoming loose. Based on the true value of the loose position in the target variable, the corresponding first category representation vector is obtained from the category deviation vector, and the category deviation vectors other than the first category representation vector in the category deviation vector are determined as the second category representation vector; A first similarity score is obtained between the first feature vector and the first category representation vector, and a second similarity score is obtained between the first feature vector and each of the second category representation vectors; The deviation between the first similarity score and the preset similarity score threshold is determined to obtain the first deviation value; Based on the first deviation value and each of the second similarity scores, the learned first cost parameter is obtained; The first deviation value is inversely proportional to the first cost parameter, and the second similarity score is directly proportional to the first cost parameter.

2. The method according to claim 1, characterized in that, The tower bolt loosening location model includes a first intermediate network structure and a second intermediate network structure; the second abnormal soundprint instance is loaded into the tower bolt loosening location model, and the tower bolt loosening location in the sample tower bolt in the second abnormal soundprint instance is inferred through the tower bolt loosening location model to obtain a first inference result, including: The second abnormal voiceprint instance is loaded into the first intermediate network structure, and a feature extraction operation is performed on the second abnormal voiceprint instance through the first intermediate network structure to obtain the second feature vector. The second feature vector is loaded into the second intermediate network structure, and a scalar multiplication operation is performed on the second feature vector using the category bias parameter through the second intermediate network structure to obtain the fourth element sequence; wherein, the category bias parameter includes multiple category bias vectors, each category bias vector corresponds to a preset loosening position, and the number of elements in the fourth element sequence is the same as the number of preset loosening positions; The fourth element sequence is normalized by performing a localization activation function to obtain the fifth element sequence, and the fifth element sequence is determined as the first inference result; wherein, the elements in the fifth element sequence represent the first confidence value of the preset loose position of the sample tower bolt inferred by the tower bolt loose position localization model.

3. The method according to claim 2, characterized in that, The obtained second cost parameter includes: The third similarity score between the second feature vector and each of the category deviation vectors is obtained; Based on the location confidence threshold, a similarity score threshold is obtained; The deviation between the third similarity score with the maximum value and the similarity score threshold is determined to obtain the third deviation value; Based on the third deviation value, the learned second cost parameter is obtained; The third deviation value is directly proportional to the second cost parameter.

4. The method according to claim 1, characterized in that, The process of acquiring the feedback acoustic signature data of the target tower bolt during the overall monitoring cycle and calculating the structural damage risk coefficient of the tower structure where the target tower bolt is located includes: Obtain the original feedback acoustic data of the target tower bolts during the monitoring period; Perform data denoising on the original feedback acoustic fingerprint data of the target tower bolt during the monitoring period to obtain the denoised feedback acoustic fingerprint data for the monitoring period. Perform a data dimensionality reduction operation on the denoised feedback voiceprint data to obtain the dimensionality-reduced feedback voiceprint data for the monitoring period. Perform data normalization operations on the dimension-reduced feedback voiceprint data for the monitoring period to obtain the feedback voiceprint data for the monitoring period; The feedback voiceprint data from multiple monitoring periods are used to construct the feedback voiceprint data for the overall monitoring cycle, wherein the overall monitoring cycle consists of multiple monitoring periods, and each monitoring period includes multiple voiceprint acquisition time nodes. Obtain the initial trend coefficient of the feedback voiceprint data during the monitoring period and the initial baseline value of the feedback voiceprint data during the monitoring period; Based on the initial baseline value and the initial trend coefficient of the feedback voiceprint data, a first conversion of the voiceprint acquisition time node is performed to obtain the first conversion output. When the error value between the first conversion output and the feedback voiceprint data at the voiceprint acquisition time node is lower than the tolerance range threshold, the partial fitting error of the quadratic growth relationship between the voiceprint acquisition time node and the error value is obtained. When the error value between the first conversion output and the feedback voiceprint data at the voiceprint acquisition time node is greater than the tolerance range threshold, a first benchmark value with a quadratic growth relationship with the tolerance range threshold is obtained, the error value is multiplied by the tolerance range threshold, and the time trend coefficient is subtracted from the first benchmark value to obtain the voiceprint acquisition time node partial fitting error. The partial fitting error is obtained by summing the partial fitting errors of the multiple voiceprint acquisition time nodes. With the goal of reducing the partial fitting error, the initial baseline value of the feedback voiceprint data and the initial trend coefficient of the feedback voiceprint data are updated to obtain the first baseline value of the feedback voiceprint data and the first trend coefficient of the feedback voiceprint data. The first trend coefficient of the feedback voiceprint data, the first baseline value of the feedback voiceprint data, and the first difference of the feedback voiceprint data constitute the first voiceprint feature parameter of the feedback voiceprint data. Based on the first voiceprint feature parameters, the target voiceprint acquisition time node of the monitoring period is converted to obtain partial voiceprint trend data of the monitoring period. An integration operation is performed on partial voiceprint trend data from multiple monitoring periods to obtain overall voiceprint trend data. Then, univariate regression processing is performed on the overall voiceprint trend data to obtain the second trend coefficient of the feedback voiceprint data and the second baseline value of the feedback voiceprint data. The second trend coefficient of the feedback voiceprint data, the second baseline value of the feedback voiceprint data, and the second error of the feedback voiceprint data constitute the second voiceprint feature parameter of the feedback voiceprint data. Data conversion is performed based on the second voiceprint feature parameters to obtain the overall voiceprint trend data for the overall monitoring period. The overall acoustic signature trend data of the target tower bolt is converted into confidence data to obtain the structural damage risk coefficient of the tower structure where the target tower bolt is located.

5. The method according to claim 4, characterized in that, The acquisition of feedback acoustic signature data of the target tower bolts during the overall monitoring cycle includes: The frequency characteristics, amplitude characteristics, and time-domain characteristics of the target tower bolt at the acoustic signature acquisition time node are obtained. The frequency characteristics, amplitude characteristics, and time domain characteristics of the target tower bolt at the acoustic signature acquisition time node constitute the original feedback acoustic signature data of the target tower bolt at the acoustic signature acquisition time node. The original feedback acoustic data of the target tower bolt at multiple acoustic data acquisition time points constitute the original feedback acoustic data of the target tower bolt during the monitoring period; the original feedback acoustic data of the target tower bolt during the monitoring period includes the original feedback acoustic data of the target tower bolt at each acoustic data acquisition time point during the monitoring period. The original feedback acoustic data of the target tower bolt at each acoustic data acquisition time node in the monitoring period that are missing acoustic data and acoustic noise data are screened out to obtain the denoised original feedback acoustic data. Redundant data elimination is performed on the original feedback voiceprint data of each voiceprint acquisition time node in the denoised original feedback voiceprint data to obtain the denoised feedback voiceprint data of the monitoring period. Perform a data dimensionality reduction operation on the denoised feedback voiceprint data to obtain the dimensionality-reduced feedback voiceprint data for the monitoring period; the dimensionality-reduced feedback voiceprint data for the monitoring period includes the dimensionality-reduced feedback voiceprint data for each voiceprint acquisition time node in the monitoring period in multiple voiceprint feature dimensions; Obtain the maximum and minimum dimensionality-reduced feedback voiceprint data among the dimensionality-reduced feedback voiceprint data of the voiceprint feature dimension at multiple voiceprint acquisition time nodes. Obtain the first difference value between the maximum dimension-reduced feedback voiceprint data and the minimum dimension-reduced feedback voiceprint data; Obtain a second difference value between the dimensionality-reduced feedback voiceprint data of the voiceprint feature dimension at each of the voiceprint acquisition time nodes and the minimum dimensionality-reduced feedback voiceprint data; The ratio between the second difference value and the first difference value corresponding to each of the voiceprint acquisition time nodes is used as the standardized feedback voiceprint data of the voiceprint acquisition time node in the voiceprint feature dimension. The standardized feedback voiceprint data of the voiceprint acquisition time node at multiple voiceprint feature dimensions are used to construct the feedback voiceprint data of the voiceprint acquisition time node. The feedback voiceprint data from multiple voiceprint acquisition time points constitutes the feedback voiceprint data for the monitoring period. The feedback voiceprint data from multiple monitoring periods are used to construct the feedback voiceprint data for the overall monitoring cycle.

6. The method according to claim 4, characterized in that, The target voiceprint acquisition time nodes include the first voiceprint acquisition time node and the last voiceprint acquisition time node. The data conversion of the target voiceprint acquisition time nodes for the monitoring period based on the first voiceprint feature parameters yields partial voiceprint trend data for the monitoring period, including: The first trend coefficient in the first voiceprint feature parameter is multiplied by the first voiceprint acquisition time node in the monitoring period to obtain the first time trend coefficient. The first time trend coefficient is added to the first baseline value to obtain partial voiceprint trend data corresponding to the first voiceprint acquisition time node. The first trend coefficient in the first voiceprint feature parameter is multiplied by the last voiceprint acquisition time node in the monitoring period to obtain the second time trend coefficient. The second time trend coefficient is added to the first baseline value to obtain partial voiceprint trend data corresponding to the last voiceprint acquisition time node; The partial voiceprint trend data for the monitoring period is composed of the partial voiceprint trend data corresponding to the first voiceprint acquisition time node and the partial voiceprint trend data corresponding to the last voiceprint acquisition time node.

7. The method according to claim 4, characterized in that, Before performing confidence data transformation on the overall acoustic signature trend data of the target tower bolt to obtain the structural damage risk coefficient of the target tower bolt, the method further includes: The feedback acoustic data of the entire monitoring period before the actual tower structure was damaged was obtained as the target sample instance, and the feedback acoustic data of the entire monitoring period after the actual tower structure was repaired was obtained as the non-target sample instance. Obtain the overall voiceprint trend data corresponding to the target sample instance and the overall voiceprint trend data corresponding to the non-target sample instance; The numerical region of the target sample instance is generated based on the overall voiceprint trend data of the target sample instance, and the numerical region of the non-target sample instance is generated based on the overall voiceprint trend data of the non-target sample instance. The overlapping portion between the numerical region of the target sample instance and the numerical region of the non-target sample instance is defined as the risk numerical region. The process of converting the overall acoustic signature trend data of the target tower bolts into confidence data to obtain the structural damage risk coefficient of the target tower bolts includes: When the overall acoustic signature trend data of the target tower bolt is within the risk value range, the confidence level corresponding to the risk value range is used as the structural damage risk coefficient of the tower structure where the target tower bolt is located.

8. A server system, characterized in that, Includes a server, said server being used to perform the method of any one of claims 1-7.

Images