Tower bolt health detection method and system based on artificial intelligence and voiceprint

Abstract

The present invention discloses a tower bolt health detection method and system based on artificial intelligence and voiceprint, comprising: first, obtaining pending feedback voiceprint instance data containing true values, and these true values ​​represent the corresponding tower bolt health status. Afterwards, these data are balanced and voiceprint features are extracted therefrom. Next, a tower bolt health status classification model is trained based on the extracted voiceprint features. Finally, the current feedback voiceprint set is collected, its voiceprint features are extracted, and the trained model is used to identify the tower bolt health status. Through the predicted true value, the tower bolt health status corresponding to the current feedback voiceprint can be understood. With such a design, the automatic detection and identification of the tower bolt health status is realized through artificial intelligence technology, which improves the detection efficiency and reduces the misjudgment rate.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a tower bolt health detection method and system based on artificial intelligence and voiceprint. Background Art

[0002] In the daily maintenance of the tower, it is crucial to monitor the health of the bolts.

[0003] For example, a Chinese patent with publication number CN116839883A discloses a method and device for diagnosing loose screws on a tower based on voiceprint recognition, which relates to the field of fault diagnosis technology. The real-time tower information of the target tower is obtained, and the target sound wave signal generated by the sound wave generator is adaptively determined according to the real-time tower information; the target sound wave signal is emitted to the target tower, and the soundprint signals at each screw connection of the target tower are collected; for the soundprint signal at each screw connection, the MFCC feature vector of the soundprint signal is extracted, and the MFCC feature vector is weighted and optimized to obtain the target feature vector; the target feature vector is input into the preset acoustic model to obtain the diagnosis and identification result of whether the screw is loose. The adaptive active sound system is adopted, and the sound content is automatically analyzed and adapted according to the tower model, size, environment and other conditions, which greatly optimizes the efficiency of soundprint analysis, and the preset acoustic model is used to efficiently scan and analyze the tightening status of the bolts of the entire tower, determine the looseness of the bolts, and improve the detection accuracy and work efficiency.

[0004] For example, the Chinese patent with publication number CN112786059A discloses a voiceprint feature extraction method and device based on artificial intelligence. The present invention includes the following steps: collecting voice data and non-voice data to establish a sample database; taking an audio file from the sample database and processing it to obtain an audio frame sequence; performing Fourier transformation on each frame in the audio frame sequence to obtain the spectrum diagram information corresponding to the frame; extracting time domain information and frequency domain information to obtain time domain features and frequency domain features; performing feature aggregation on time domain features and frequency domain features to obtain aggregated features; performing vector embedding on aggregated features to obtain voiceprint feature vectors; inputting the voiceprint feature vectors into a convolutional neural network model for training to obtain a voiceprint feature model; obtaining the voice data to be recognized for preprocessing; and inputting the preprocessed voice feature data into the voiceprint feature model to obtain a voice feature vector. The present invention improves the accuracy and efficiency of voiceprint feature extraction.

[0005] Another example is a Chinese patent with publication number CN111076960B, which discloses a voiceprint quality detection method based on an artificial intelligence algorithm, including a voiceprint storage module and a marking module in a PC; the implementation method includes the following steps: the first step: firstly, the voiceprint features of known air conditioner faults are labeled by the marking module, and these data are divided into a training set and a test set according to a certain ratio; the second step: design a suitable artificial intelligence deep learning type for the features of the air conditioner voiceprint; the third step: train the model with the labeled voiceprint data; the fourth step: test the empirical model with the test set to see if it is correct; the fifth step: input the voiceprint to be detected into the model, and the model gives the data for detecting the voiceprint; the sixth step: continuously optimize the empirical model. The present invention can still accurately find various quality problems such as air conditioners under interference and changeable conditions, covering the four major types of fault problem identification of air conditioners, and designs a reasonable software architecture so that the system has good expansion and adaptive learning capabilities.

[0006] Currently, the main way to detect the health status of tower bolts is through manual visual inspection and manual testing, and the existing technologies focus on the technical point of screw loosening, and do not mention the health monitoring of bolts. However, these traditional methods are not only inefficient, but also may affect the accuracy of the detection results due to human factors. Summary of the invention

[0007] The purpose of the present invention is to provide a tower bolt health detection method and system based on artificial intelligence and voiceprint.

[0008] In a first aspect, an embodiment of the present invention provides a tower bolt health detection method based on artificial intelligence and voiceprint, comprising:

[0009] Acquire pending feedback voiceprint instance data, wherein the pending feedback voiceprint instance data includes a tower bolt feedback voiceprint bound to a real value, wherein the real value represents a health state of the tower bolt corresponding to the tower bolt feedback voiceprint;

[0010] Performing data balancing on the pending feedback voiceprint instance data to obtain a feedback voiceprint observation set, wherein the number of positive voiceprint instances in the feedback voiceprint observation set is greater than the number of negative voiceprint instances;

[0011] Performing a feature extraction operation on the feedback voiceprint observation set to obtain a voiceprint feature of each feedback voiceprint instance in the feedback voiceprint observation set;

[0012] Training a tower bolt health status classification model according to the feedback voiceprint instances in the feedback voiceprint observation set and the voiceprint features of each feedback voiceprint instance, to obtain a trained tower bolt health status classification model;

[0013] Performing a feature extraction operation on the collected current feedback voiceprint set to obtain a voiceprint feature of each current feedback voiceprint in the current feedback voiceprint set;

[0014] The health status of the tower bolts is identified based on the trained tower bolt health status classification model and the voiceprint features of each current feedback voiceprint in the current feedback voiceprint set, and the predicted true value of each current feedback voiceprint in the current feedback voiceprint set is obtained, and the predicted true value represents the health status of the tower bolt corresponding to the current feedback voiceprint.

[0015] In a second aspect, an embodiment of the present invention provides a server system, including a server, wherein the server is used to execute the method described in the first aspect.

[0016] Compared with the prior art, the beneficial effects provided by the present invention include: adopting the tower bolt health detection method and system based on artificial intelligence and voiceprint disclosed in the present invention, by obtaining pending feedback voiceprint instance data containing true values, these true values ​​represent the corresponding tower bolt health status. Afterwards, these data are balanced and voiceprint features are extracted from them. Then, a tower bolt health status classification model is trained based on the extracted voiceprint features. Finally, the current feedback voiceprint set is collected, its voiceprint features are extracted, and the trained model is used to identify the tower bolt health status. Through the predicted true value, the tower bolt health status corresponding to the current feedback voiceprint can be understood. Such a design not only improves the detection efficiency, but also significantly improves the accuracy of the detection results. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can also be obtained based on these drawings without creative work.

[0018] Figure 1 A schematic diagram of the steps of a tower bolt health detection method based on artificial intelligence and voiceprint provided in an embodiment of the present invention;

[0019] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0021] The specific implementation modes of the present invention are described in detail below in conjunction with the accompanying drawings.

[0022] In order to solve the technical problems in the aforementioned background technology, Figure 1 A flow chart of a tower bolt health detection method based on artificial intelligence and voiceprint provided in an embodiment of the present disclosure is provided. The following is a detailed introduction to the tower bolt health detection method based on artificial intelligence and voiceprint.

[0023] Step S201, obtaining pending feedback voiceprint instance data, wherein the pending feedback voiceprint instance data includes a tower bolt feedback voiceprint bound to a real value, wherein the real value represents a health state of the tower bolt corresponding to the tower bolt feedback voiceprint;

[0024] In the embodiment of the present invention, the above steps are exemplary to obtain voiceprint data that has not been processed or classified from a data source (such as a sensor, a database, etc.). These data are raw, unprocessed, and require further analysis to determine their meaning or status. For example, there is a tower maintenance system equipped with a sound sensor for capturing the sound emitted by the tower bolts. The server regularly collects sound data from these sensors, which are the so-called "pending feedback voiceprint instance data". These data have not been analyzed, so it is impossible to directly know the health status of the bolts they correspond to. "Pending" means that these data have not been assigned any labels or classifications, and their meanings are unknown or to be determined. "Feedback voiceprint instance data" refers to sound samples collected from a specific object (in this case, the tower bolts), which can be used for subsequent analysis and pattern recognition. For example, the server collects a series of sound files from the sound sensor, each of which contains the sound emitted by the tower bolts over a period of time. These sound files are "pending feedback voiceprint instance data", because at this stage, it is not known what kind of health status of the bolts these sounds represent. "Binding to the real value" means that each voiceprint instance is associated with a known actual state (i.e., the real value). This real value is determined by other means (such as manual inspection, other sensor data, etc.) and is considered accurate. "Tower bolt feedback voiceprint" refers to the sound samples collected from the tower bolts, which reflect the current state of the bolts. During the tower maintenance process, the staff can regularly inspect the bolts and record their health status (such as normal, loose, damaged, etc.). At the same time, the sound sensor is constantly collecting the sounds emitted by these bolts. The server matches these sound data with the real state recorded by the staff to form a labeled data set. In this way, each sound sample (i.e., feedback voiceprint) is "bound" to a corresponding real value (such as "normal" or "damaged"). "Real value" refers to the actual health state of the tower bolt determined by a reliable method. "Representation" means that this real value can represent or describe the bolt state corresponding to the voiceprint data. In other words, there is a corresponding relationship between the voiceprint data and the real value, which can be used for subsequent pattern recognition and prediction. For example, the server analyzes the sound data and finds that a specific voiceprint pattern is highly correlated with the state of the bolt loosening. When this voiceprint pattern appears again, the server can predict that the corresponding bolt may be loose. This prediction result needs to be compared with the "true value" to verify its accuracy. If the true value also shows that the bolt is loose, then the server's prediction is accurate, and this voiceprint pattern effectively "characterizes" the state of the bolt being loose.

[0025] Step S202, performing data balancing on the pending feedback voiceprint instance data to obtain a feedback voiceprint observation set, wherein the number of positive voiceprint instances in the feedback voiceprint observation set is greater than the number of negative voiceprint instances;

[0026] In the embodiment of the present invention, the pending feedback voiceprint instance data refers to the collected original voiceprint data that has not been processed or classified. These data contain the sound information emitted by the tower bolts in different health states, but at this stage, their corresponding health states are not known. For example, there is a database that stores a large number of sound files collected from the tower bolts. These files are "pending feedback voiceprint instance data" because at this stage, no processing or analysis has been performed on these data. Data balancing is a data processing technology used to solve the problem of unbalanced number of samples of different categories in the data set. In machine learning, if the number of samples of a certain category is much larger than that of other categories, the model may over-focus on the category with a large number and ignore the category with a small number during the training process. The purpose of data balancing is to make the number of samples of different categories relatively balanced, thereby improving the generalization ability of the model. In the tower bolt health detection scenario, the voiceprint data in the "normal" state may be much larger than the data in the "damaged" state. In order to balance the number of these two types of data, some strategies can be adopted, such as oversampling the minority class (increasing the amount of data in the "damaged" state) or undersampling the majority class (reducing the amount of data in the "normal" state). After this processing, a relatively balanced data set can be obtained. The data set obtained after data balancing is the "feedback voiceprint observation set". The voiceprint data in this set has been preprocessed, and the number of samples of different categories is relatively balanced, which can be used for subsequent model training. Continuing the above example, after data balancing, a set of voiceprint data in "normal" and "damaged" states is obtained, in which the number of data in the two categories is relatively close. This set is the "feedback voiceprint observation set". In the binary classification problem of machine learning, the category of interest (such as disease, fault, etc.) is usually marked as positive, and other categories are marked as negative. In this scenario, "positive voiceprint instance" can represent the voiceprint data of the bolt in an unhealthy state (such as loose, damaged), and "negative voiceprint instance" can represent the voiceprint data of the bolt in a healthy state. This sentence means that in the feedback voiceprint observation set obtained after data balancing, the number of positive voiceprint instances is greater than the number of negative voiceprint instances.

[0027] Step S203, performing a feature extraction operation on the feedback voiceprint observation set to obtain a voiceprint feature of each feedback voiceprint instance in the feedback voiceprint observation set;

[0028] In an embodiment of the present invention, exemplarily, a "feedback voiceprint observation set" refers to a voiceprint data set that has been processed by data balancing. This set contains multiple voiceprint instances, each of which represents the sound information collected from the tower bolts within a specific period of time. These sound information have been preprocessed, and the number of samples in different categories is relatively balanced, making them more suitable for training machine learning models. For example, there are 1,000 voiceprint instances, 500 of which are the sounds of bolts in a normal state, and the other 500 are the sounds of bolts in a damaged state. These 1,000 instances together constitute the "feedback voiceprint observation set". "Feature extraction" is a key step in machine learning and data processing. It refers to extracting meaningful and representative information or attributes from raw data. These information or attributes can reflect the essential characteristics of the data and can be used for subsequent model training and prediction. In voiceprint processing, feature extraction usually involves the conversion and analysis of sound signals to extract key parameters that can describe the characteristics of the sound. For voiceprint data, common feature extraction methods include extracting basic information such as frequency, amplitude, and duration of the sound, or more complex features such as Mel Frequency Cepstral Coefficients (MFCC) and Linear Predictive Coding (LPC). These features can help distinguish the sounds of bolts in different states. For example, a damaged bolt can emit specific frequency components or abnormal amplitude changes, which can be captured and quantified through feature extraction operations. "Voiceprint features of feedback voiceprint instances" refers to the feature values ​​or feature vectors extracted from each specific voiceprint instance. These features describe the characteristics of the sound signal represented by the instance and can be used as input for machine learning models. For example, a feature extraction operation is performed on a specific voiceprint instance to obtain a feature vector containing 10 values. This vector can contain information such as the frequency distribution, energy distribution, and time domain characteristics of the sound. This feature vector is the "voiceprint feature" of the voiceprint instance, which can be input into the machine learning model for training and predicting the health status of the bolt. In the scenario of tower bolt health detection, the server will first perform data balancing on the collected raw voiceprint data to obtain a feedback voiceprint observation set. The server then performs feature extraction on each voiceprint instance in this set, extracting key parameters or feature vectors that can describe the sound characteristics. These extracted voiceprint features will be used in subsequent machine learning model training and prediction processes. By analyzing and comparing these features, the health status of tower bolts can be more accurately identified and predicted.

[0029] Step S204, training a tower bolt health status classification model according to the feedback voiceprint instances in the feedback voiceprint observation set and the voiceprint features of each feedback voiceprint instance, to obtain a trained tower bolt health status classification model;

[0030] In the embodiment of the present invention, the "feedback soundprint observation set" is a soundprint data set after data balancing, which contains multiple feedback soundprint instances, which represent the sound information emitted by the tower bolts in different health states. This set is the basic data set for machine learning model training. For example, the database stores 1,000 balanced soundprint instances, of which 500 are the sounds of bolts in normal state and 500 are the sounds of bolts in damaged state. These 1,000 instances constitute the "feedback soundprint observation set". "Feedback soundprint instance" refers to a specific, single soundprint data sample. Each instance contains the sound information emitted by the tower bolts at a specific time and has been preprocessed and feature extracted. In the feedback soundprint observation set, each individual sound file or sound record can be regarded as a "feedback soundprint instance". For example, a sound collected from a tower bolt at a specific time point is a feedback soundprint instance. "Voiceprint feature" refers to the key parameter or feature vector extracted from the feedback soundprint instance that can describe the sound characteristics. These features are used as input to the machine learning model to help the model learn and identify bolt sounds in different health states. For each feedback soundprint instance, its frequency distribution, energy distribution, time domain characteristics, etc. can be extracted as soundprint features. For example, a specific frequency peak or energy change pattern can be an important soundprint feature. The "Tower Bolt Health Status Classification Model" is a machine learning model whose task is to predict and classify the health status of the tower bolts based on the input soundprint features. This model will learn how to distinguish different health states based on soundprint features during the training process. Various machine learning algorithms can be used to build this classification model, such as support vector machines (SVM), decision trees, random forests, or deep learning networks. During the training process, the model receives the soundprint features of the feedback soundprint instances as input and tries to predict the bolt health status (such as normal, damaged, etc.) corresponding to these instances. By comparing the model's prediction results with the actual health status labels, the model's parameters can be continuously adjusted to make it more accurate and reliable. The "Trained Tower Bolt Health Status Classification Model" refers to a machine learning model that has achieved a certain prediction accuracy and stability after enough training iterations. This model has learned how to identify the health status of tower bolts from voiceprint features and can be used for actual prediction tasks. Once the tower bolt health status classification model achieves a satisfactory accuracy rate (for example, more than 95%) on the training data and its performance on new data is relatively stable, it can be said that the model has been "trained". At this point, the model can be deployed to actual application scenarios, such as embedded in the tower maintenance system for real-time monitoring and early warning of the health status of bolts.

[0031] Step S205, performing a feature extraction operation on the collected current feedback voiceprint set to obtain the voiceprint feature of each current feedback voiceprint in the current feedback voiceprint set;

[0032] In the embodiment of the present invention, exemplarily, the "collected current feedback soundprint set" refers to a set of newly collected soundprint data in actual application. These data can come from the real-time monitoring of the tower bolts, representing the sound information emitted by the bolts at the current time point or time period. Unlike the previous "feedback soundprint observation set", the "current feedback soundprint set" here emphasizes real-time and up-to-dateness. For example, the health status of the bolts of a tower is being monitored, and a series of sound data are collected in real time through the sound sensor installed on the tower. These data constitute the "collected current feedback soundprint set". "Current feedback soundprint" refers to a single soundprint data sample in the current feedback soundprint set. Each current feedback soundprint represents the sound information emitted by the tower bolt at a specific moment. In the real-time monitoring process, each sound record captured by the sound sensor can be regarded as a "current feedback soundprint". For example, at a specific time point, the sound emitted by a bolt captured by the sensor is a current feedback soundprint. "Feature extraction operation" refers to the process of processing and analyzing the soundprint data to extract key parameters or feature vectors that can describe the sound characteristics. These features will be used as input for subsequent machine learning models to help the model predict and classify the health status of the bolts. For each current feedback voiceprint, a series of signal processing and mathematical transformations can be performed to extract basic information such as frequency, amplitude, duration, or more complex features such as Mel Frequency Cepstral Coefficients (MFCC) and Linear Predictive Coding (LPC). These extracted parameters or feature vectors are the features of the voiceprint. "Voiceprint features" refer to key parameters or feature vectors that can describe the characteristics of the sound extracted from the current feedback voiceprint through feature extraction operations. These features are the basis for machine learning models to make predictions and classifications. For example, a feature extraction operation is performed on a current feedback voiceprint to obtain a feature vector containing 10 values. This vector can contain information such as the frequency distribution, energy distribution, and time domain characteristics of the sound. This feature vector is the "voiceprint feature" of the current feedback voiceprint. In actual applications, the sound data emitted by the tower bolts will be collected in real time to form a "collected current feedback voiceprint set". Then, a feature extraction operation is performed on each current feedback voiceprint to extract key parameters or feature vectors that can describe the characteristics of the sound, namely "voiceprint features". These extracted voiceprint features will be input into the previously trained tower bolt health status classification model to predict and classify the current health status of the bolts. By continuously collecting and analyzing new voiceprint data, real-time monitoring and early warning of the health status of tower bolts can be achieved.

[0033] Step S206, identifying the health status of the tower bolts according to the trained tower bolt health status classification model and the voiceprint features of each current feedback voiceprint in the current feedback voiceprint set, and obtaining the predicted true value of each current feedback voiceprint in the current feedback voiceprint set, wherein the predicted true value represents the health status of the tower bolts corresponding to the current feedback voiceprint.

[0034] In the embodiment of the present invention, the trained tower bolt health status classification model is, for example, a machine learning model that has been fully trained and has stable performance. The goal of the model is to predict the health status of the tower bolts based on the input voiceprint features. During the training process, the model learns how to distinguish the normal and abnormal states of the bolts based on the features in the historical voiceprint data. For example, a deep learning network (such as a convolutional neural network or a recurrent neural network) is used to construct this classification model. After training on a large amount of voiceprint data, the model has learned to extract key information from the voiceprint features, such as frequency mode, energy distribution, etc., to distinguish whether the bolts are normal or abnormal. This refers to the result obtained after performing a feature extraction operation on each voiceprint sample in the currently collected voiceprint data set. These features describe the sound characteristics of each voiceprint sample and are the basis for the model to perform state recognition. For each voiceprint sample in the current feedback voiceprint set, the same feature extraction method (such as MFCC, LPC, etc.) is used to extract its frequency, amplitude, duration and other features. In this way, each voiceprint sample will be converted into a feature vector for subsequent prediction and classification. This refers to the process of using the trained model to predict and classify each voiceprint sample in the current feedback voiceprint set. The model will determine whether the corresponding tower bolt is in a normal state or an abnormal state based on the characteristics of each voiceprint sample. For each voiceprint sample in the current feedback voiceprint set, its feature vector is input into the trained classification model. The model will give a prediction value based on these features, such as "normal" or "abnormal", which is the recognition result of the health status of the tower bolt. This refers to the health status of the tower bolt corresponding to each voiceprint sample predicted by the model. These prediction values ​​are estimates of the actual situation and can be used for further analysis and decision-making. If the health status of 100 voiceprint samples in the current feedback voiceprint set is identified, the model can give 80 "normal" and 20 "abnormal" prediction results. These prediction results are the predicted true values ​​of each voiceprint sample, which reflect the model's judgment on the health status of the current tower bolt. This means that the predicted true value is not only an explanation of the model output, but also a reflection of the actual health status of the tower bolt. The predicted true value can help understand which bolts may have problems, so that corresponding maintenance and repair measures can be taken. If the model predicts that the bolt status of a voiceprint sample is "abnormal", it can be considered that there may be problems with the tower bolts and further inspection and repair are required. In this way, the predicted true value provides direct information about the health status of the tower bolts.

[0035] In order to more clearly describe the solution provided by the embodiments of the present application, an overall implementation method is described below.

[0036] In an embodiment of the present invention, illustratively, the server receives and stores a large amount of feedback soundprint data of tower bolts from multiple sensors. These data are collected in an actual tower environment and include the sounds emitted by bolts in different health states. Each piece of data is associated with a true value, which represents the health state of the bolt, such as "normal", "loose" or "damaged". The server processes the collected data and finds that the amount of data in some health states is significantly more than that in other states, for example, the data in the "normal" state may be much more than the data in the "damaged" state. In order to balance the amount of data in different categories, the server uses data enhancement technology (such as adding noise, time domain stretching, etc.) to increase the number of samples in the minority class, thereby constructing a feedback soundprint observation set, in which the number of instances of positive class soundprints (such as the "damaged" state) is relatively balanced with the number of instances of negative class soundprints (such as the "normal" state). The server performs feature extraction on the balanced feedback soundprint observation set. It uses audio processing algorithms such as Fourier transform and Mel-Frequency Cepstral Coefficients (MFCC) to extract key sound features such as frequency, amplitude, timbre, etc. from each voiceprint instance, which can reflect the health status of the bolt. The server uses the extracted voiceprint features and the corresponding true values ​​to train a tower bolt health status classification model. This model can be a deep learning model, such as a convolutional neural network (CNN) or a recurrent neural network (RNN), which can predict the health status of the bolt based on the input voiceprint features. During the training process, the server continuously adjusts the parameters of the model to minimize the difference between the predicted value and the true value. The server receives the current feedback voiceprint data from the sensor on the tower in real time. It immediately processes this data and extracts the same voiceprint features as when training the model. These features will be used for subsequent bolt health status identification. The server inputs the extracted voiceprint features of the current feedback voiceprint into the previously trained tower bolt health status classification model. The model predicts based on these features and outputs the predicted true value corresponding to each current feedback voiceprint. These predicted true values ​​represent the current health status of the bolt, such as "normal", "needs attention" or "urgent repair", etc. The server sends this information to maintenance personnel in real time so that they can take corresponding measures in time.

[0037] In the embodiment of the present invention, the aforementioned step S205 can be implemented through the following example execution.

[0038] (1) obtaining the bolt acoustic features of each current feedback voiceprint in the current feedback voiceprint set;

[0039] (2) obtaining a bolt fault type feature of each current feedback voiceprint in the current feedback voiceprint set according to a bolt health assessment indicator of each current feedback voiceprint in the current feedback voiceprint set;

[0040] (3) performing a feature fusion operation on the bolt acoustic features and bolt fault type features of each current feedback soundprint in the current feedback soundprint set in combination with a preset detection time sequence, to obtain the bolt acoustic features after feature fusion and the bolt fault type features after feature fusion at different detection time sequences;

[0041] (4) performing a feature extraction operation on the bolt acoustic features and bolt fault type features before and after the feature fusion to obtain a first bolt state voiceprint representation of each current feedback voiceprint in the current feedback voiceprint set;

[0042] (5) processing the state classification feature in the first bolt state voiceprint representation of each current feedback voiceprint in the current feedback voiceprint set to obtain the second bolt state voiceprint representation of each current feedback voiceprint in the current feedback voiceprint set;

[0043] (6) Integrate the first bolt state voiceprint representation and the second bolt state voiceprint representation of each current feedback voiceprint in the current feedback voiceprint set to obtain the voiceprint feature of each current feedback voiceprint in the current feedback voiceprint set.

[0044] In an embodiment of the present invention, illustratively, the server receives a set of voiceprint data transmitted in real time from a sound sensor installed on an iron tower, and these data constitute the current feedback voiceprint set. The server first preprocesses these original sound data, such as denoising, standardization, etc. Subsequently, the server uses a specific signal processing algorithm, such as Fourier transform or wavelet analysis, to extract the acoustic features of the bolts in each voiceprint sample, such as frequency distribution, amplitude change, resonance peak, etc. The server analyzes each voiceprint sample in the current feedback voiceprint set according to a pre-set bolt health assessment indicator, such as a specific frequency component or energy distribution pattern in the sound. By comparing these indicators with a voiceprint feature library of known fault types, the server can determine the bolt fault type that each voiceprint sample may correspond to, such as looseness, fracture or corrosion, and extract the corresponding fault type features. The server fuses the bolt acoustic features and fault type features extracted in step 1 and step 2 according to a preset detection sequence, such as multiple voiceprint sampling points arranged in chronological order. This fusion may include weighted averaging, feature splicing, or feature fusion through a deep learning model. The fused features will contain both time series information and fault type information, which helps to more accurately describe the health status of the bolts. The server performs further feature extraction operations on the acoustic features and fault type features of the bolts before and after fusion. This can include dimensionality reduction processing (such as principal component analysis PCA), feature selection (such as methods based on statistical tests) or feature transformation (such as kernel methods). Through these operations, the server can extract more representative and discriminative features for subsequent bolt status identification. The server constructs the first bolt status voiceprint representation of each voiceprint sample based on the extracted fusion features, and processes the status classification features in these representations. The processing operations may include normalization, discretization, or preliminary classification using specific classification algorithms (such as support vector machines SVM, random forests, etc.). These processing operations help improve the accuracy and stability of subsequent bolt status identification. Finally, the server integrates the first bolt status voiceprint representation (based on the features before fusion) and the second bolt status voiceprint representation (based on the features after fusion) of each voiceprint sample. The integration operation may include feature concatenation, weighted averaging, or feature fusion through a deep learning model. The integrated voiceprint features will contain information of both original features and fused features, which will help describe the health status of the bolt more comprehensively. These integrated voiceprint features will be used in subsequent bolt health status classification and prediction tasks.

[0045] In the embodiment of the present invention, the aforementioned step of obtaining the bolt acoustic features of each current feedback voiceprint in the current feedback voiceprint set can be implemented through the following example.

[0046] (1) Acquiring, according to the bolt voiceprint history data of each current feedback voiceprint in the current feedback voiceprint set, the bolt acoustic features of each current feedback voiceprint in the current feedback voiceprint set, wherein the bolt acoustic features include bolt basic features and voiceprint collection device parameters.

[0047] In the embodiment of the present invention, the server stores a large amount of historical voiceprint data, which includes the sound information emitted by the tower bolts at different times and in different environments. When the server receives new voiceprint data, that is, the current feedback voiceprint set, it will first access these historical data. By comparing the newly collected voiceprint with the historical voiceprint data, the server can identify the historical pattern that is most similar to the current voiceprint.

[0048] These historical voiceprint data contain the basic characteristics of the bolts, such as frequency, amplitude, phase, etc. These basic characteristics are inherent properties of the sound signal, and they will be different for different types of bolts and different health states. The server uses advanced signal processing technologies, such as Fourier analysis and wavelet transform, to extract these basic characteristics of the bolts from the current feedback voiceprint.

[0049] In addition, the parameters of the voiceprint collection device also have an important impact on the extracted voiceprint features. Different sensors, different sampling rates, different environmental noises, etc. will affect the final collected voiceprint data. Therefore, when the server extracts the acoustic features of the bolt, it will also take into account the parameters of these voiceprint collection devices. For example, if the frequency response characteristics of a certain sensor are known, the server can make corresponding corrections to the collected voiceprint data to eliminate the errors introduced by the sensor.

[0050] Based on the above information, the server can accurately extract the acoustic features of the bolts of each current feedback soundprint in the current feedback soundprint set. These features not only include the sound characteristics of the bolts themselves, but also reflect the influence of the collection equipment and environmental factors, providing a comprehensive and accurate data basis for the subsequent bolt health status assessment.

[0051] In the embodiment of the present invention, the aforementioned step of performing feature extraction operation on the bolt acoustic features and bolt fault type features before and after feature fusion can be implemented through the following example.

[0052] (1) performing a feature scaling operation on the continuous variables in the bolt acoustic features and the bolt fault type features before and after the feature fusion;

[0053] (2) Performing feature quantization operations on the categorical variables in the bolt acoustic features and bolt fault type features before and after the feature fusion.

[0054] In an embodiment of the present invention, exemplarily, when processing the acoustic features of the bolts and the features of the bolt fault types before and after feature fusion, the server first identifies the continuous variables therein. Continuous variables, such as the frequency and amplitude of sound, have continuously changing values. In order to eliminate the dimensional differences and numerical range differences between different features, the server performs feature scaling operations on these continuous variables. For example, the server can use the Z-score normalization method to convert the value of each continuous variable into a standard normal distribution with a mean of 0 and a standard deviation of 1. In this way, the values ​​of different features are in the same order of magnitude, which helps to improve the training efficiency and accuracy of subsequent machine learning models. In addition to continuous variables, the feature set can also include categorical variables, such as the fault type of the bolt (loose, corroded, broken, etc.). These categorical variables usually exist in the form of text or discrete values ​​and cannot be directly used for numerical calculations. Therefore, the server needs to perform feature quantization operations on these categorical variables to convert them into numerical data. A common feature quantization method is One-Hot Encoding. The server creates a new binary feature column for each different value of the categorical variable. For example, for the categorical variable of bolt failure type, the server can create three new feature columns: "loose", "corrosion" and "fracture", and the value of each feature column can only be 0 or 1. Then, according to the value of the original categorical variable, the server will fill in 1 in the corresponding feature column and 0 in the remaining feature columns. In this way, the original categorical variable is converted into a numerical binary feature vector, which can be used for subsequent numerical calculations and model training.

[0055] In an embodiment of the present invention, the aforementioned step of integrating the first bolt state voiceprint representation and the second bolt state voiceprint representation of each current feedback voiceprint in the current feedback voiceprint set to obtain the voiceprint features of each current feedback voiceprint in the current feedback voiceprint set can be implemented through the following example.

[0056] (1) The low-order voiceprint features and the high-order voiceprint features in the first bolt state voiceprint representation and the second bolt state voiceprint representation of each current feedback voiceprint in the current feedback voiceprint set are concatenated in columns to obtain the voiceprint features of each current feedback voiceprint in the current feedback voiceprint set.

[0057] In the embodiment of the present invention, illustratively, after the server completes the extraction of the first bolt state voiceprint representation and the second bolt state voiceprint representation of each voiceprint in the current feedback voiceprint set, it enters the feature integration stage. In this stage, the server needs to combine different types of voiceprint features to form a more comprehensive and more distinguishing feature set.

[0058] First, the server will identify the low-order voiceprint features and high-order voiceprint features in the voiceprint representation of the first bolt state and the voiceprint representation of the second bolt state. Low-order voiceprint features usually refer to basic features extracted directly from the original voiceprint data, such as time domain waveform features, frequency domain spectrum features, etc. These features are more intuitive, but can contain more redundant information and noise. High-order voiceprint features are abstract features extracted from low-order features through more complex mathematical transformations or machine learning algorithms, such as principal components after principal component analysis (PCA), hidden layer outputs in deep learning models, etc. These features are more discriminative and robust, but relatively less explanatory.

[0059] Then, the server will combine the low-order voiceprint features and high-order voiceprint features in the first bolt state voiceprint representation and the second bolt state voiceprint representation by column splicing. Specifically, it connects the feature vectors of different types end to end to form a longer feature vector. In this way, each current feedback voiceprint corresponds to a voiceprint feature vector containing rich information.

[0060] By integrating low-order and high-order voiceprint features, the server can make full use of the complementarity between different types of features to improve the accuracy and stability of subsequent bolt status recognition tasks. At the same time, this integration method also helps to reduce feature dimensions and computational complexity, improving overall processing efficiency.

[0061] In the embodiment of the present invention, illustratively, the aforementioned step S201 may be implemented through the following example execution.

[0062] (1) According to the voiceprint screening criteria and health monitoring logic, the tower bolt feedback voiceprint bound to the real value is obtained from the original feedback voiceprint instance data, wherein the voiceprint screening criteria includes voiceprint data source verification, and the health monitoring logic includes the relationship between the bolt position and the historical health status;

[0063] (2) obtaining bolt acoustic information of the tower bolt feedback soundprint;

[0064] (3) determining a fault soundprint identification index based on the bolt acoustic information of the tower bolt feedback soundprint;

[0065] (4) Based on the fault voiceprint identification index and the voiceprint abnormal distribution judgment criterion, the abnormal tower bolt feedback voiceprint is screened out from the tower bolt feedback voiceprint to obtain the optimized tower bolt feedback voiceprint bound to the true value.

[0066] In an embodiment of the present invention, exemplarily, the server receives a large amount of original feedback voiceprint instance data, which comes from different tower bolts and contains various environmental noises and interferences. In order to ensure the accuracy and reliability of the data, the server first filters the data according to the voiceprint screening criteria. One of the voiceprint screening criteria is voiceprint data source verification, that is, confirming whether the voiceprint data comes from a known and reliable tower bolt monitoring point. At the same time, the server also uses health monitoring logic to further filter the data. This includes considering the relationship between the bolt position and the historical health status. For example, if the bolts at a certain position have frequently failed in history, then the new voiceprint data at that position will be regarded as a key focus. By comprehensively applying the voiceprint screening criteria and health monitoring logic, the server can filter out the tower bolt feedback voiceprints bound to the real value from the original feedback voiceprint instance data, that is, those voiceprint data associated with the actual bolt health status. After filtering out the tower bolt feedback voiceprints bound to the real value, the server further extracts the bolt acoustic information in these voiceprints. This includes acoustic features such as frequency, amplitude, and phase, which can reflect the health status and potential fault types of the bolts. The server uses advanced signal processing technology to analyze and process these soundprints and extract key bolt acoustic information. This information will be used in the subsequent fault soundprint identification index determination and abnormal soundprint screening process. The server determines the indicators for identifying fault soundprints based on the extracted bolt acoustic information. These indicators can be specific frequency thresholds, amplitude change rates, or phase differences, which can indicate whether the bolt is faulty and the type and degree of the fault. By comparing and analyzing the acoustic information of healthy bolts and faulty bolts, the server can establish an effective fault soundprint identification indicator system. This indicator system will provide an important basis for subsequent soundprint abnormality judgment and fault identification. After determining the fault soundprint identification indicators, the server begins to apply these indicators and the soundprint abnormal distribution judgment criteria to screen out abnormal tower bolt feedback soundprints. The soundprint abnormal distribution judgment criteria can include the statistical characteristics of the soundprint data, the change rules of the time domain or frequency domain characteristics, etc. The server will evaluate each tower bolt feedback soundprint to determine whether it meets the requirements of the fault soundprint identification indicators and whether there is an abnormal distribution. Those voiceprints that do not meet the requirements or have abnormalities will be regarded as abnormal tower bolt feedback voiceprints and screened out from the data set. Finally, the server obtains the optimized tower bolt feedback voiceprint data set bound to the real value, which will be used for subsequent model training and fault identification tasks.

[0067] In the embodiment of the present invention, the aforementioned step S202 may be implemented through the following example execution.

[0068] (1) performing data balancing on the pending feedback voiceprint instance data according to the intelligent sampling strategy of the minority feedback voiceprint instance to obtain a feedback voiceprint observation set; or

[0069] (2) Performing data balancing on the pending feedback voiceprint instance data according to a deep learning strategy to obtain a feedback voiceprint observation set.

[0070] In the embodiment of the present invention, for example, in actual applications, the pending feedback voiceprint instance data collected by the server may have a class imbalance problem, that is, the number of voiceprint instances in some categories is much larger than that in other categories. This imbalance will affect the training effect and performance of the subsequent machine learning model. In order to solve this problem, the server needs to balance the data. The server first analyzes the class distribution of the pending feedback voiceprint instance data and finds that the number of minority class feedback voiceprint instances (such as a specific type of fault voiceprint) is much smaller than that of other categories. In order to balance the amount of data between different categories, the server uses an intelligent sampling strategy to enhance the minority class feedback voiceprint instances. The intelligent sampling strategy can include two methods: oversampling and undersampling. Oversampling is to increase the number of minority class samples by copying or generating new minority class samples, while undersampling is to achieve balance by reducing the number of majority class samples. The server can select a suitable sampling method according to the actual situation. In the intelligent sampling process, the server can also use some advanced machine learning technologies, such as synthetic minority class oversampling technology (SMOTE) or adaptive synthetic sampling (ADASYN), to generate new minority class samples or select the most representative majority class samples for undersampling. By applying intelligent sampling strategies, the server can effectively balance the amount of data between different categories and obtain a more balanced set of feedback voiceprint observations. In addition to intelligent sampling strategies, the server can also use deep learning strategies to balance the pending feedback voiceprint instance data. Deep learning models have powerful feature extraction and representation learning capabilities. They can generate new samples or improve the model's recognition ability for minority classes by learning the features of minority class samples. A common deep learning strategy is to use a generative adversarial network (GAN) to generate new minority class samples. GAN consists of two parts: a generator and a discriminator. The generator is responsible for generating new samples, while the discriminator is responsible for judging whether the generated samples are real. By training the GAN model, the server can generate new samples that are similar to real minority class samples, thereby increasing the amount of minority class data. In addition, the server can also use deep learning strategies such as transfer learning, data enhancement, or class weight adjustment to further improve the model's recognition performance for minority classes. Transfer learning can use the knowledge learned in other related tasks to assist the training of the current task; data enhancement can increase the diversity of samples by rotating, cropping, adding noise and other operations to the original samples; category weight adjustment can give a larger weight to minority samples during the training process, so that they have a greater impact on model training. Through the application of deep learning strategies, the server can make full use of the powerful capabilities of deep learning models to balance the pending feedback voiceprint instance data and obtain a more balanced and high-quality feedback voiceprint observation set.

[0071] In the embodiment of the present invention, the step of performing data balancing on the pending feedback voiceprint instance data according to the intelligent sampling strategy of the minority class feedback voiceprint instance to obtain the feedback voiceprint observation set can be implemented through the following example.

[0072] (1) Obtaining the pending feedback voiceprint instance data;

[0073] (2) performing cluster analysis on the majority of the feedback voiceprint instance data in the pending feedback voiceprint instance data;

[0074] (3) performing data balancing on the pending feedback voiceprint instance data after the cluster analysis according to the ensemble learning strategy to obtain a transition data set;

[0075] (4) Determining the transition data set as the feedback voiceprint observation set and inputting it into the self-attention mechanism model for processing, and obtaining the confidence that each feedback voiceprint instance in the feedback voiceprint observation set is classified as a positive voiceprint instance;

[0076] (5) selecting a fuzzy feedback voiceprint instance based on the maximum entropy principle and the confidence that each feedback voiceprint instance in the feedback voiceprint observation set is classified as a positive voiceprint instance;

[0077] (6) Placing the selected fuzzy feedback voiceprint instance into the feedback voiceprint observation set to obtain an optimized feedback voiceprint observation set.

[0078] In an embodiment of the present invention, exemplarily, the server retrieves pending feedback voiceprint instance data from the voiceprint database. These data may come from different monitoring points and contain information on the health status of various bolts. The server loads these data into the memory and prepares for subsequent data balancing processing. The server performs category analysis on the pending feedback voiceprint instance data loaded into the memory and finds that there are a large number of majority class feedback voiceprint instances (such as bolt voiceprints in normal state). In order to process these data more effectively, the server uses cluster analysis technology to divide the majority class feedback voiceprint instances into different clusters. The voiceprint instances in each cluster have similar characteristics, which is convenient for subsequent data balancing processing. The server performs sampling processing on the majority class feedback voiceprint instances in each cluster according to the ensemble learning strategy. The ensemble learning strategy combines a variety of sampling methods, such as random undersampling, clustering-based undersampling, etc., to ensure that the representative characteristics are retained while reducing the number of majority class samples. After sampling processing, the server obtains a transition data set that is relatively balanced in number with the minority class feedback voiceprint instances. The server takes the transition data set as input and passes it to a self-attention mechanism model for processing. The self-attention mechanism model can automatically learn the correlation and importance between voiceprint instances, and generate a confidence score for each voiceprint instance that is classified as a positive class (such as a fault state). This score reflects the degree of certainty of the model's judgment that the voiceprint instance belongs to the positive class. According to the maximum entropy principle, the server selects those voiceprint instances that are classified as positive class voiceprint instances with low confidence (i.e., fuzzy) from the feedback voiceprint observation set. The maximum entropy principle tends to select samples with the greatest uncertainty for further processing to improve the generalization ability of the model. These selected fuzzy feedback voiceprint instances will become the focus of subsequent data balancing. The server will relocate the selected fuzzy feedback voiceprint instances to the feedback voiceprint observation set and adjust their weights or quantities to ensure that they receive sufficient attention in the subsequent training process. The feedback voiceprint observation set processed in this way will have better balance and representativeness, which will help improve the recognition performance of the machine learning model for various voiceprint instances.

[0079] In the embodiment of the present invention, the aforementioned step of performing data balancing on the pending feedback voiceprint instance data according to the deep learning strategy to obtain the feedback voiceprint observation set can be implemented through the following example.

[0080] (1) Obtaining the pending feedback voiceprint instance data;

[0081] (2) extracting high-order voiceprint features of each feedback voiceprint instance in the pending feedback voiceprint instance data;

[0082] (3) inputting the high-order voiceprint features of each feedback voiceprint instance in the pending feedback voiceprint instance data into the self-attention mechanism model for processing, and obtaining the confidence that each feedback voiceprint instance in the pending feedback voiceprint instance data is classified as a positive voiceprint instance;

[0083] (4) based on the confidence level of each feedback voiceprint instance in the pending feedback voiceprint instance data being classified as a positive voiceprint instance, a preset number of feedback voiceprint instances with higher confidence levels are selected from the pending feedback voiceprint instance data as key feedback voiceprint instances;

[0084] (5) retraining the self-attention mechanism model based on the original feedback voiceprint instance in the pending feedback voiceprint instance data and the determined key feedback voiceprint instance;

[0085] (6) Re-determine the key feedback voiceprint instances by completing the training of the self-attention mechanism model, so that the number of positive voiceprint instances in the obtained feedback voiceprint observation set is greater than the number of negative voiceprint instances.

[0086] In an embodiment of the present invention, illustratively, the server retrieves pending feedback voiceprint instance data from the voiceprint database. These data contain bolt voiceprint information collected from various environments, including both normal voiceprints and voiceprints that may have faults. The server loads these data into the memory and prepares for subsequent feature extraction and model processing. The server uses the feature extraction method in deep learning technology to process the pending feedback voiceprint instance data loaded into the memory. Through model structures such as convolutional neural networks (CNN) or recurrent neural networks (RNN), the server extracts high-order voiceprint features for each feedback voiceprint instance. These high-order features can more accurately describe the complex patterns and potential structures of voiceprints, which helps to improve the accuracy of subsequent classification. The server uses the extracted high-order voiceprint features as input and passes them to a self-attention mechanism model for processing. The self-attention mechanism model can automatically learn the correlation and importance between voiceprint features, and generate a confidence score for each feedback voiceprint instance classified as a positive class (such as a fault state). This score reflects the degree of certainty of the model's judgment that the voiceprint instance belongs to the positive class. Based on the generated confidence scores, the server selects a preset number of voiceprint instances with high confidence values ​​from the pending feedback voiceprint instance data as key feedback voiceprint instances. These key instances are considered to be more likely to belong to the positive class (such as fault status), and therefore will receive more attention in subsequent data balancing and model training. The server merges the original feedback voiceprint instances and the determined key feedback voiceprint instances to form a new training data set. Then, the self-attention mechanism model is trained twice using this data set. During the training process, the model will pay more attention to those voiceprint instances marked as key, thereby optimizing its recognition performance for positive voiceprint instances. After the second training, the self-attention mechanism model has improved its recognition ability for positive voiceprint instances. The server uses the trained model again to process the pending feedback voiceprint instance data and re-determines the key feedback voiceprint instances. The key instances determined this time will more accurately reflect the characteristics and distribution of the positive voiceprint instances. Finally, the server adds these key instances to the feedback voiceprint observation set to ensure that the number of positive voiceprint instances is greater than the number of negative voiceprint instances, thereby achieving data balancing.

[0087] In the embodiment of the present invention, the aforementioned step S203 can be implemented through the following example execution.

[0088] (1) obtaining the bolt acoustic features of each feedback voiceprint instance in the feedback voiceprint observation set;

[0089] (2) obtaining bolt fault type characteristics of each feedback voiceprint instance in the feedback voiceprint observation set according to the bolt health assessment index of each feedback voiceprint instance in the feedback voiceprint observation set;

[0090] (3) performing a feature fusion operation on the bolt acoustic features and bolt fault type features of each feedback soundprint instance in the feedback soundprint observation set in combination with a preset detection time sequence, to obtain the bolt acoustic features after feature fusion and the bolt fault type features after feature fusion at different detection time sequences;

[0091] (4) performing a feature extraction operation on the bolt acoustic features and bolt fault type features before and after the feature fusion to obtain a first bolt state voiceprint representation of each feedback voiceprint instance in the feedback voiceprint observation set;

[0092] (5) processing the state classification features in the first bolt state voiceprint representation of each feedback voiceprint instance in the feedback voiceprint observation set to obtain the second bolt state voiceprint representation of each feedback voiceprint instance in the feedback voiceprint observation set;

[0093] (6) Performing an integration operation on the first bolt state voiceprint representation and the second bolt state voiceprint representation of each feedback voiceprint instance in the feedback voiceprint observation set to obtain the voiceprint feature of each feedback voiceprint instance in the feedback voiceprint observation set.

[0094] In the embodiment of the present invention, exemplarily, after receiving the feedback soundprint observation set, the server first extracts the acoustic features of the bolt for each feedback soundprint instance. These acoustic features may include acoustic parameters related to time such as frequency, amplitude, and phase, which reflect the vibration and sound characteristics of the bolt in a specific working environment. The server uses advanced signal processing algorithms to extract and quantify these acoustic parameters to form a bolt acoustic feature vector for each feedback soundprint instance. In addition to the acoustic features, the server also obtains its fault type features according to the bolt health assessment indicators corresponding to each feedback soundprint instance. These health assessment indicators may include the tightening force, looseness, wear, etc. of the bolt, which are closely related to the fault type of the bolt. The server determines the possible fault type corresponding to each feedback soundprint instance by analyzing these health assessment indicators, and encodes it into a fault type feature vector. Considering that the importance of the bolt acoustic features and the fault type features may be different at different detection time sequences, the server performs a fusion operation on these features in combination with the preset detection time sequence. For example, in the early detection time sequence, the acoustic features may be more important; while in the later detection time sequence, the fault type features may be more decisive. According to this time sequence dependency, the server performs weighted fusion of acoustic features and fault type features to obtain fused feature vectors under different detection time sequences. Based on feature fusion, the server further performs deep feature extraction operations on the acoustic features and fault type features before fusion and the features after fusion. This may involve using deep learning models (such as convolutional neural networks or recurrent neural networks) to automatically learn and extract higher-level abstract features. Through these operations, the server obtains the first bolt state voiceprint representation of each feedback voiceprint instance, which more comprehensively reflects the health status and fault type information of the bolt. The first bolt state voiceprint representation contains rich state classification features, such as the tightness and wear degree of the bolt. The server further processes and analyzes these state classification features to extract more representative feature information. This may involve using feature selection algorithms (such as principal component analysis or mutual information method) to filter out the most important features, or using feature conversion methods (such as feature hashing or embedding representation) to convert the original features into a form that is easier for the model to learn. Through these processing operations, the server obtains the second bolt state voiceprint representation of each feedback voiceprint instance. Finally, the server integrates the first bolt state voiceprint representation and the second bolt state voiceprint representation of each feedback voiceprint instance. This may involve concatenating, weighted summing, or other forms of fusion of the two representation vectors to form a more complete and comprehensive voiceprint feature vector. This voiceprint feature vector not only contains the original acoustic information and fault type information, but also incorporates higher-level abstract features and representative features obtained through deep learning and feature processing. These voiceprint features will serve as important inputs for subsequent model training and prediction.

[0095] In the embodiment of the present invention, the aforementioned step of obtaining the bolt acoustic features of each feedback voiceprint instance in the feedback voiceprint observation set can be implemented through the following example.

[0096] (1) According to the bolt voiceprint history data of each feedback voiceprint instance in the feedback voiceprint observation set, the bolt acoustic features of each feedback voiceprint instance in the feedback voiceprint observation set are obtained, wherein the bolt acoustic features include bolt basic features and voiceprint collection equipment parameters.

[0097] In the embodiment of the present invention, for example, a large amount of bolt voiceprint history data is stored in the storage system of the server. These data are obtained by monitoring the bolts through the voiceprint collection device at different time points and environments. When the server needs to obtain the bolt acoustic features of each feedback voiceprint instance in the feedback voiceprint observation set, it will perform the following operations:

[0098] First, the server retrieves the bolt voiceprint history data corresponding to each feedback voiceprint instance from the storage system. These data may include information such as the sound waveform, frequency spectrum, amplitude, etc. of the bolt under different working conditions, as well as the parameter settings used by the voiceprint collection device when collecting these sound data, such as sampling rate, resolution, microphone type, etc. Next, the server processes and analyzes these bolt voiceprint history data. It uses advanced signal processing techniques, such as Fourier transform and wavelet analysis, to extract key acoustic features in the bolt sound, such as the intensity of specific frequency components, the time domain and frequency domain characteristics of the sound, etc. These acoustic features can reflect the working state and performance of the bolt. At the same time, the server also considers the impact of the parameters of the voiceprint collection device on the acoustic features. Different voiceprint collection devices can have different sensitivities, frequency responses, and noise characteristics, all of which will affect the collected sound data. Therefore, the server will also extract and record the parameters of the voiceprint collection device as part of the bolt acoustic features. Finally, the server integrates the extracted basic acoustic features of the bolt with the parameters of the voiceprint collection device to form a complete bolt acoustic feature vector for each feedback voiceprint instance. These feature vectors will be used in subsequent model training and prediction tasks to improve the ability to identify and evaluate the working status of bolts.

[0099] In the embodiment of the present invention, the aforementioned step of performing feature extraction operation on the bolt acoustic features and bolt fault type features before and after feature fusion can be implemented through the following example.

[0100] (1) performing a feature scaling operation on the continuous variables in the bolt acoustic features and the bolt fault type features before and after the feature fusion;

[0101] (2) Performing feature quantization operations on the categorical variables in the bolt acoustic features and bolt fault type features before and after the feature fusion.

[0102] In an embodiment of the present invention, exemplarily, in the data processing flow of the server, when extracting features of the bolt acoustic features and the bolt fault type features, different types of variables will be encountered. Among them, continuous variables are common, and they represent variables that can take any real value within a certain range, such as temperature, length, frequency, etc. For continuous variables, the server will perform feature scaling operations to ensure that these variables have similar scales in terms of values. This is because many machine learning algorithms are sensitive to the scale of features when processing features. If the numerical range of a feature is much larger than other features, it can dominate the model, resulting in a decrease in the model's attention to other features. In order to perform feature scaling, the server can use methods such as standardization or normalization. For example, standardization is to subtract the mean of the feature value and then divide it by its standard deviation so that the processed feature value has zero mean and unit variance. Normalization is to scale the feature value to a specified range (such as 0 to 1). When performing the feature scaling operation, the server will traverse each continuous variable, calculate its mean and standard deviation (or maximum and minimum values), and then convert the value of each variable according to the selected scaling method. In this way, all continuous variables will be scaled to a similar scale, providing more balanced feature inputs for subsequent model training. In addition to continuous variables, bolt acoustic features and bolt fault type features can also contain categorical variables. These variables represent discrete, incomparable category information, such as bolt model, fault type, etc. For categorical variables, the server performs feature quantization operations to convert them into numerical data so that machine learning algorithms can process them. There are many methods for feature quantization, such as one-hot encoding and label encoding. One-hot encoding is a commonly used feature quantization method that converts each categorical variable into a binary vector containing only one "1" and multiple "0s". The length of the vector is equal to the number of categories, and each category corresponds to a position in the vector. For example, if there is a categorical variable representing the bolt model, it has three possible values: A, B, and C. Then, the server generates three binary vectors for these three categories: [1,0,0], [0,1,0], and [0,0,1]. Then, according to the category value of each instance, it is converted into the corresponding binary vector. Label encoding is another simple feature quantization method that directly maps each category to an integer value. For example, for the bolt model classification variable above, the server can map it to: A->1, B->2, C->3. Then, according to the category value of each instance, it is replaced with the corresponding integer value. When performing feature quantization operations, the server will traverse each classification variable and generate a corresponding numerical representation for each variable based on its value range and the selected quantization method. In this way, all classification variables will be converted to numerical data, providing more unified and usable feature input for subsequent model training.

[0103] In an embodiment of the present invention, the aforementioned step of integrating the first bolt state voiceprint representation and the second bolt state voiceprint representation of each feedback voiceprint instance in the feedback voiceprint observation set to obtain the voiceprint features of each feedback voiceprint instance in the feedback voiceprint observation set can be implemented through the following example.

[0104] (1) The low-order voiceprint features and the high-order voiceprint features in the first bolt state voiceprint representation and the second bolt state voiceprint representation of each feedback voiceprint instance in the feedback voiceprint observation set are concatenated in columns to obtain the voiceprint features of each feedback voiceprint instance in the feedback voiceprint observation set.

[0105] In an embodiment of the present invention, exemplarily, in the data processing and analysis process of the server, after the first bolt state voiceprint representation and the second bolt state voiceprint representation of each feedback voiceprint instance have been obtained, the two representations need to be integrated to form a more comprehensive and representative voiceprint feature vector. The first bolt state voiceprint representation and the second bolt state voiceprint representation respectively contain bolt state information extracted from different angles and levels. Among them, the low-order voiceprint features mainly reflect the basic properties of the bolt sound and the directly observed features, such as frequency, amplitude, etc.; while the high-order voiceprint features are further abstracted and extracted from the low-order features through complex models such as deep learning, and they may represent deeper and more indirect information about the bolt state. In order to effectively integrate these two types of features, the server will adopt a column-by-column splicing method. Specifically, the server will first determine which dimensions and attributes the low-order voiceprint features and high-order voiceprint features in the first bolt state voiceprint representation and the second bolt state voiceprint representation correspond to. Then, it will arrange these features according to their order and position in the original representation to form a longer feature vector. In this process, the server needs to ensure that the concatenated feature vectors are consistent in dimension and value so that subsequent model training and prediction tasks can proceed smoothly. If the feature dimensions in the first bolt state voiceprint representation and the second bolt state voiceprint representation do not match or the numerical ranges are very different, the server can use some additional processing steps, such as feature selection, dimensionality reduction, or feature scaling, to ensure that the integrated voiceprint feature vector has the appropriate dimension and numerical range. Ultimately, by integrating the low-order voiceprint features and high-order voiceprint features in the first bolt state voiceprint representation and the second bolt state voiceprint representation, the server can obtain a more comprehensive and representative voiceprint feature vector, which will serve as one of the important inputs for subsequent model training and prediction.

[0106] In the embodiment of the present invention, the second bolt state voiceprint representation includes a representation vector, and the representation vector includes a tower bolt acoustic trajectory representation vector and an environmental noise representation vector.

[0107] In an embodiment of the present invention, exemplarily, in the voiceprint processing flow of the server, when acoustic monitoring of the tower bolts is performed, a large amount of sound data will be collected. These data not only contain the sound of the bolts themselves, but may also be interfered by the surrounding environmental noise. In order to more accurately analyze and identify the state of the bolts, the server needs to carefully process and characterize these sound data. The second bolt state voiceprint characterization is an important link in this processing flow. It aims to extract key information from complex sound data and represent the acoustic characteristics of the bolts and the characteristics of the environmental noise in the form of vectors. The server will first pre-process the collected sound data, such as filtering, noise reduction, etc., to highlight the sound signal of the bolts. Then, through advanced signal processing techniques, such as Fourier transform, wavelet analysis, etc., the server will extract the key features of the bolt sound, such as frequency, amplitude, phase, etc. These features can reflect the working state and performance of the bolts. Subsequently, the server encodes and integrates these features to form a tower bolt acoustic trajectory representation vector. This vector depicts the changing trajectory of the bolt sound in the time or frequency domain in numerical form, providing important input for subsequent model training and prediction. In addition to the sound of bolts, the sound data can also contain a lot of environmental noise, such as wind noise, mechanical operation noise, etc. These noises will interfere with the recognition and analysis of bolt sounds. To solve this problem, the server will extract and characterize the environmental noise in the sound data separately. Through noise recognition algorithms and feature extraction techniques, the server can identify the environmental noise components in the sound data and extract its key features, such as noise type and intensity. Then, the server encodes and integrates these features to form an environmental noise representation vector. This vector describes the characteristics and intensity of environmental noise in numerical form, which helps the server to effectively suppress or filter noise in subsequent processing. In summary, the second bolt state voiceprint representation provides a comprehensive and accurate feature description of bolt sound and environmental noise by combining the tower bolt acoustic trajectory representation vector and the environmental noise representation vector. This provides strong support for subsequent tasks such as bolt state recognition and fault diagnosis.

[0108] In the embodiment of the present invention, the aforementioned step S204 may be implemented through the following example execution.

[0109] (1) Dividing the feedback voiceprint observation set after feature engineering into feedback voiceprint training data and feedback voiceprint test data;

[0110] (2) training a plurality of potential models to be determined simultaneously according to the original model parameters and the feedback voiceprint training data, so as to determine a model with the best model performance from the plurality of potential models to be determined as the tower bolt health status classification model;

[0111] (3) performing hyperparameter optimization on the tower bolt health status classification model;

[0112] (4) According to the model parameters after hyperparameter optimization, the tower bolt health status classification model is trained for the second time to obtain a trained tower bolt health status classification model, and a verification operation is performed on the trained tower bolt health status classification model using a plurality of the feedback voiceprint verification data.

[0113] In an embodiment of the present invention, illustratively, after receiving the feedback voiceprint observation set that has been processed by feature engineering, the server will first divide the data. This set contains a large number of feedback voiceprint instances, each of which is accompanied by its corresponding voiceprint features. In order to effectively train the model and evaluate its performance, the server needs to divide this set into two parts: one part is used to train the model, and the other part is used to test the model. The server will use methods such as random sampling or stratified sampling to ensure the consistency of the distribution of training data and test data. For example, the server can divide the feedback voiceprint observation set into feedback voiceprint training data and feedback voiceprint test data in a ratio of 70% and 30%. In this way, the training data contains enough information to train the model, and the test data can be used to evaluate the performance of the model on unseen data. After dividing the data, the server will start training the model. It will initialize and train multiple potential models to be determined at the same time based on the preset original model parameters (such as the number of layers and nodes of the neural network) and the feedback voiceprint training data. These models can have different structures or parameter settings to explore different model spaces and find the optimal model configuration. During the training process, the server uses the input features (i.e., voiceprint features) and corresponding labels (i.e., bolt health status) in the feedback voiceprint training data to train each model. It continuously adjusts the model's parameters through optimization techniques such as the back-propagation algorithm to minimize the prediction error and improve the model's classification performance. After training, the server uses the feedback voiceprint test data to evaluate the performance of each model. It calculates the accuracy, recall, F1 score and other indicators of each model on the test data, and compares these indicators to determine the model with the best performance. This optimal model will be selected as the tower bolt health status classification model. After selecting the optimal model, the server will further optimize its hyperparameters. Hyperparameters are parameters set before model training. They have an important impact on the performance of the model, but cannot be directly learned through the training process. The server uses methods such as grid search, random search or Bayesian optimization to explore different hyperparameter combinations. For example, it can adjust the values ​​of hyperparameters such as learning rate, batch size, regularization strength, and observe the impact of these changes on model performance. Through multiple attempts and comparisons, the server can find a set of optimal hyperparameter configurations that achieve the best performance on the test data. After determining the optimal hyperparameter configuration, the server will use these configurations to perform a second training on the tower bolt health status classification model. This training will make full use of the information in the feedback voiceprint training data and combine it with the optimal hyperparameter settings to further improve the performance of the model. After the second training is completed, the server will obtain a fully optimized and trained tower bolt health status classification model. In order to verify the performance and generalization ability of this model, the server will use multiple feedback voiceprint verification data to verify it.These verification data are voiceprint data collected from actual application scenarios and are independent of the training data. During the verification process, the server will input the verification data into the model and observe the degree of match between the model's output and the actual label. It will calculate the model's accuracy, recall and other indicators on the verification data, and evaluate whether the model's performance meets the needs of actual applications. If the model's performance meets the expected standards, the server can deploy it to the actual system for real-time monitoring and classification of bolt health status.

[0114] In an embodiment of the present invention, after the aforementioned step of obtaining a trained tower bolt health status classification model, the embodiment of the present invention also provides the following implementation method.

[0115] (1) locking the trained tower bolt health status classification model;

[0116] (2) Locking the training process of the trained tower bolt health status classification model.

[0117] In the embodiment of the present invention, for example, after the server completes the training of the tower bolt health status classification model and verifies that its performance meets the requirements, it will perform a model locking operation. The purpose of this step is to ensure that the model maintains its state and performance stability in the subsequent application process and prevent unauthorized modification or misoperation from damaging the model.

[0118] Specifically, the server will use specific software tools or platform functions to lock the trained model files. The locking operation can include setting the file's read-only permission, adding digital signatures or checksums, and other measures to ensure the integrity and credibility of the model files. In this way, even if other users or programs try to modify the model files, they will fail due to insufficient permissions or verification failure.

[0119] In addition, the server can also store locked model files in a secure and controllable environment, such as an encrypted file system or a dedicated model repository. Only authenticated and authorized users can access and use these model files, further ensuring the security and stability of the model. In addition to locking the model itself, the server also locks the model training process. The purpose of this step is to preserve the historical information and context of model training so that the model training process and results can be reproduced when needed.

[0120] Specifically, the server will record all key steps and operations in the model training process, including data preprocessing, model structure design, hyperparameter adjustment, training algorithm selection, etc. These records can be saved and managed in the form of log files, configuration files, or version control tools. At the same time, the server will also save the data set, label information, and evaluation indicators used in the training process for subsequent analysis and comparison.

[0121] By locking the model's training process, the server can not only reproduce the model's training results when needed, but also continuously monitor and improve the model's performance and stability. If problems with the model are found in actual applications or further optimization is required, the server can locate the problem and formulate a solution based on the locked training process and data.

[0122] In short, locking the trained tower bolt health status classification model and its training process is one of the important measures to ensure the stability and credibility of the model. Through locking operations and continuous management and monitoring, the server can ensure that the model plays a stable role in subsequent application processes and provide reliable support for tower bolt health status monitoring and classification.

[0123] The embodiment of the present 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 above-mentioned tower bolt health detection method based on artificial intelligence and voiceprint. Figure 2 As shown, Figure 2 The block diagram of the computer device 100 provided in the embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112 and a communication unit 113. To achieve data transmission or interaction, the memory 111, the processor 112 and the communication unit 113 are directly or indirectly electrically connected to each other. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

[0124] For illustrative purposes, the foregoing description is made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise form disclosed. Numerous modifications and variations are possible in accordance with the above teachings. These embodiments are selected and described in order to best illustrate the principles of the present disclosure and its practical application, so that those skilled in the art can best utilize the present disclosure and utilize various embodiments with different modifications to suit the intended specific application.

Claims

1. A tower bolt health detection method based on artificial intelligence and voiceprint, characterized in that: include: Acquire pending feedback voiceprint instance data, wherein the pending feedback voiceprint instance data includes a tower bolt feedback voiceprint bound to a real value, wherein the real value represents a health state of the tower bolt corresponding to the tower bolt feedback voiceprint; Performing data balancing on the pending feedback voiceprint instance data to obtain a feedback voiceprint observation set, wherein the number of positive voiceprint instances in the feedback voiceprint observation set is greater than the number of negative voiceprint instances; Performing a feature extraction operation on the feedback voiceprint observation set to obtain a voiceprint feature of each feedback voiceprint instance in the feedback voiceprint observation set; Training a tower bolt health status classification model according to the feedback voiceprint instances in the feedback voiceprint observation set and the voiceprint features of each feedback voiceprint instance, to obtain a trained tower bolt health status classification model; Performing a feature extraction operation on the collected current feedback voiceprint set to obtain a voiceprint feature of each current feedback voiceprint in the current feedback voiceprint set; The health status of the tower bolts is identified according to the trained tower bolt health status classification model and the voiceprint features of each current feedback voiceprint in the current feedback voiceprint set, and the predicted true value of each current feedback voiceprint in the current feedback voiceprint set is obtained, and the predicted true value represents the health status of the tower bolts corresponding to the current feedback voiceprint. The step of performing data balancing on the pending feedback voiceprint instance data to obtain a feedback voiceprint observation set includes: Acquire the pending feedback voiceprint instance data; Performing cluster analysis on the majority of the feedback voiceprint instance data in the pending feedback voiceprint instance data; According to the ensemble learning strategy, data balancing is performed on the pending feedback voiceprint instance data after the cluster analysis to obtain a transition data set; Determine the transition data set as the feedback voiceprint observation set and input it into the self-attention mechanism model for processing, and obtain the confidence that each feedback voiceprint instance in the feedback voiceprint observation set is classified as a positive voiceprint instance; Selecting a fuzzy feedback voiceprint instance based on the maximum entropy principle and the confidence that each feedback voiceprint instance in the feedback voiceprint observation set is classified as a positive voiceprint instance; placing the selected fuzzy feedback voiceprint instance into the feedback voiceprint observation set to obtain an optimized feedback voiceprint observation set; or Acquire the pending feedback voiceprint instance data; Extracting high-order voiceprint features of each feedback voiceprint instance in the pending feedback voiceprint instance data; Inputting the high-order voiceprint features of each feedback voiceprint instance in the pending feedback voiceprint instance data into the self-attention mechanism model for processing, and obtaining the confidence that each feedback voiceprint instance in the pending feedback voiceprint instance data is classified as a positive voiceprint instance; Based on the confidence that each feedback voiceprint instance in the pending feedback voiceprint instance data is classified as a positive voiceprint instance, a preset number of feedback voiceprint instances with top confidence values ​​are selected from the pending feedback voiceprint instance data as key feedback voiceprint instances; Based on the original feedback voiceprint instance in the pending feedback voiceprint instance data and the determined key feedback voiceprint instance, retraining the self-attention mechanism model; By completing the training of the self-attention mechanism model, the key feedback voiceprint instances are re-determined so that the number of positive voiceprint instances in the obtained feedback voiceprint observation set is greater than the number of negative voiceprint instances.

2. The method according to claim 1, characterized in that The performing a feature extraction operation on the collected current feedback voiceprint set to obtain the voiceprint feature of each current feedback voiceprint in the current feedback voiceprint set includes: According to the bolt voiceprint history data of each current feedback voiceprint in the current feedback voiceprint set, the bolt acoustic feature of each current feedback voiceprint in the current feedback voiceprint set is obtained, wherein the bolt acoustic feature includes the bolt basic feature and the voiceprint collection device parameter; According to the bolt health assessment index of each current feedback voiceprint in the current feedback voiceprint set, obtaining the bolt fault type feature of each current feedback voiceprint in the current feedback voiceprint set; Performing a feature fusion operation on the bolt acoustic features and bolt fault type features of each current feedback soundprint in the current feedback soundprint set in combination with a preset detection time sequence, to obtain the bolt acoustic features after feature fusion and the bolt fault type features after feature fusion at different detection time sequences; Performing a feature scaling operation on the continuous variables in the bolt acoustic features and the bolt fault type features before and after the feature fusion; Performing a feature quantization operation on the classification variables in the bolt acoustic features and bolt fault type features before and after the feature fusion to obtain a first bolt state voiceprint representation of each current feedback voiceprint in the current feedback voiceprint set; Processing the state classification feature in the first bolt state voiceprint representation of each current feedback voiceprint in the current feedback voiceprint set to obtain the second bolt state voiceprint representation of each current feedback voiceprint in the current feedback voiceprint set; An integration operation is performed on the first bolt state voiceprint representation and the second bolt state voiceprint representation of each current feedback voiceprint in the current feedback voiceprint set to obtain the voiceprint feature of each current feedback voiceprint in the current feedback voiceprint set.

3. The method according to claim 2, characterized in that The step of integrating the first bolt state voiceprint representation and the second bolt state voiceprint representation of each current feedback voiceprint in the current feedback voiceprint set to obtain the voiceprint feature of each current feedback voiceprint in the current feedback voiceprint set includes: The low-order voiceprint features and the high-order voiceprint features in the first bolt state voiceprint representation and the second bolt state voiceprint representation of each current feedback voiceprint in the current feedback voiceprint set are spliced ​​in columns to obtain the voiceprint features of each current feedback voiceprint in the current feedback voiceprint set.

4. The method according to claim 1, characterized in that: The obtaining of pending feedback voiceprint instance data includes: According to the voiceprint screening criteria and health monitoring logic, the tower bolt feedback voiceprint bound to the real value is obtained from the original feedback voiceprint instance data, the voiceprint screening criteria includes voiceprint data source verification, and the health monitoring logic includes the relationship between the bolt position and the historical health status; Acquiring bolt acoustic information of the tower bolt feedback soundprint; Determining a fault soundprint recognition index based on the bolt acoustic information of the tower bolt feedback soundprint; Based on the fault voiceprint identification index and the voiceprint abnormal distribution judgment criterion, abnormal tower bolt feedback voiceprints are screened out from the tower bolt feedback voiceprints to obtain optimized tower bolt feedback voiceprints bound to real values.

5. The method according to claim 1, characterized in that The performing a feature extraction operation on the feedback voiceprint observation set to obtain the voiceprint feature of each feedback voiceprint instance in the feedback voiceprint observation set includes: According to the bolt voiceprint history data of each feedback voiceprint instance in the feedback voiceprint observation set, the bolt acoustic features of each feedback voiceprint instance in the feedback voiceprint observation set are obtained, wherein the bolt acoustic features include bolt basic features and voiceprint collection equipment parameters; According to the bolt health assessment index of each feedback voiceprint instance in the feedback voiceprint observation set, obtaining the bolt fault type feature of each feedback voiceprint instance in the feedback voiceprint observation set; Performing a feature fusion operation on the bolt acoustic features and the bolt fault type features of each feedback soundprint instance in the feedback soundprint observation set in combination with a preset detection time sequence, to obtain the bolt acoustic features after feature fusion and the bolt fault type features after feature fusion at different detection time sequences; Performing a feature scaling operation on the continuous variables in the bolt acoustic features and the bolt fault type features before and after the feature fusion; Performing feature quantization operations on the classification variables in the bolt acoustic features and bolt fault type features before and after the feature fusion to obtain a first bolt state voiceprint representation of each feedback voiceprint instance in the feedback voiceprint observation set; Processing the state classification features in the first bolt state voiceprint representation of each feedback voiceprint instance in the feedback voiceprint observation set to obtain a second bolt state voiceprint representation of each feedback voiceprint instance in the feedback voiceprint observation set; The low-order voiceprint features and high-order voiceprint features in the first bolt state voiceprint representation and the second bolt state voiceprint representation of each feedback voiceprint instance in the feedback voiceprint observation set are spliced ​​in columns to obtain the voiceprint features of each feedback voiceprint instance in the feedback voiceprint observation set.

6. The method according to claim 2, characterized in that The second bolt state voiceprint representation includes a representation vector, and the representation vector includes a tower bolt acoustic trajectory representation vector and an environmental noise representation vector.

7. The method according to claim 1, characterized in that The step of training a tower bolt health status classification model according to the feedback voiceprint instances in the feedback voiceprint observation set and the voiceprint features of each feedback voiceprint instance to obtain a trained tower bolt health status classification model includes: The feedback voiceprint observation set after feature engineering is divided into feedback voiceprint training data and feedback voiceprint test data; According to the original model parameters and the feedback voiceprint training data, a plurality of potential models to be determined are trained simultaneously, so as to determine a model with the best model performance from the plurality of potential models to be determined as the tower bolt health status classification model; Performing hyperparameter optimization on the tower bolt health status classification model; According to the model parameters after hyperparameter optimization, the tower bolt health status classification model is trained for the second time to obtain a trained tower bolt health status classification model, and a verification operation is performed on the trained tower bolt health status classification model using multiple feedback voiceprint verification data.

8. The method according to claim 7, characterized in that After obtaining the trained tower bolt health status classification model, the method further includes: Locking the trained tower bolt health status classification model; The training process of the trained tower bolt health status classification model is locked.

9. A server system, characterized in that: The method comprises a server, wherein the server is used to execute the method according to any one of claims 1 to 8.

Images