Locomotive equipment maintenance method based on handheld voiceprint acquisition terminal, handheld voiceprint acquisition terminal and system

Abstract

The invention discloses a locomotive equipment maintenance method based on a handheld voiceprint collection terminal, the handheld voiceprint collection terminal and a system, and the method comprises the steps: firstly collecting the voiceprint and environmental parameters of locomotive equipment through the handheld terminal, then recognizing the voiceprint abnormality through a voiceprint abnormality type recognition model, and generating an equipment abnormality description in combination with the environmental parameters. And further deducing the abnormal condition of the locomotive part, and obtaining a corresponding maintenance scheme from a preset maintenance database. By means of the design, rapid recognition and accurate maintenance of the abnormity of the locomotive equipment are achieved, and the efficiency and accuracy of equipment maintenance are improved.

Description

Technical Field

[0001] The present invention relates to the field of sound measurement, and in particular to a locomotive equipment maintenance method based on a handheld voiceprint collection terminal, a handheld voiceprint collection terminal and a system. Background Art

[0002] With the development of industrial technology, the complexity and precision of locomotive equipment are increasing. Traditional locomotive equipment maintenance methods rely on manual inspection and experience judgment, which are inefficient and easily affected by human factors. In recent years, voiceprint recognition technology has gradually matured. By analyzing and identifying the sound generated when the equipment is running, it can effectively detect the operating status and abnormal conditions of the equipment. Therefore, applying voiceprint recognition technology to the maintenance and overhaul of locomotive equipment can greatly improve the accuracy and efficiency of maintenance. Summary of the invention

[0003] The object of the present invention is to provide a locomotive equipment maintenance method based on a handheld voiceprint collection terminal, a handheld voiceprint collection terminal and a system.

[0004] In a first aspect, an embodiment of the present invention provides a locomotive equipment maintenance method based on a handheld voiceprint collection terminal, comprising:

[0005] Obtain the current equipment voiceprint and environmental parameters collected by the handheld voiceprint collection terminal at the preset locomotive equipment maintenance location;

[0006] Inputting the current device voiceprint into a voiceprint abnormal type recognition model to obtain a current abnormal voiceprint recognition result corresponding to the current device voiceprint;

[0007] Generate a current locomotive equipment abnormality description according to the current abnormal voiceprint recognition result and the environmental parameters;

[0008] Inferring the abnormality of locomotive components according to the current locomotive equipment abnormality description;

[0009] A candidate locomotive equipment maintenance plan corresponding to the inferred locomotive component abnormality is obtained from a preset maintenance database as a current locomotive equipment maintenance plan.

[0010] In a second aspect, an embodiment of the present invention provides a handheld voiceprint collection terminal, including a microphone array and a sensor;

[0011] The microphone array and the sensor are used to collect the current device voiceprint and the environmental parameters in the first aspect respectively.

[0012] In a third aspect, an embodiment of the present invention provides a locomotive equipment maintenance system based on a handheld voiceprint collection terminal, comprising:

[0013] An acquisition module, used to acquire the current equipment voiceprint and environmental parameters collected by the handheld voiceprint collection terminal at a preset locomotive equipment maintenance location;

[0014] A processing module is used to input the current device voiceprint into a voiceprint abnormal type recognition model to obtain a current abnormal voiceprint recognition result corresponding to the current device voiceprint; generate a current locomotive device abnormality description according to the current abnormal voiceprint recognition result and the environmental parameters; and infer the abnormality of the locomotive component according to the current locomotive device abnormality description;

[0015] The maintenance module is used to obtain a candidate locomotive equipment maintenance plan corresponding to the inferred locomotive component abnormality from a preset maintenance database as a current locomotive equipment maintenance plan.

[0016] Compared with the prior art, the beneficial effects provided by the present invention include: using a locomotive equipment maintenance method based on a handheld voiceprint collection terminal, a handheld voiceprint collection terminal and a system disclosed in the present invention, including: firstly collecting the voiceprint and environmental parameters of the locomotive equipment through the handheld terminal, then using the voiceprint abnormality type recognition model to identify the voiceprint abnormality, and generating a description of the equipment abnormality in combination with the environmental parameters. Further inferring the abnormal conditions of the locomotive parts, and obtaining the corresponding maintenance plan from the preset maintenance database. Such a design realizes the rapid identification and accurate maintenance of locomotive equipment abnormalities, and improves the efficiency and accuracy of equipment maintenance. 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 flow chart of the steps of a locomotive equipment maintenance method based on a handheld voiceprint collection terminal provided by an embodiment of the present invention;

[0019] Figure 2 A schematic diagram of the structure of a handheld voiceprint collection terminal provided by an embodiment of the present invention;

[0020] Figure 3 A schematic diagram of the exploded structure of a handheld voiceprint collection terminal provided by an embodiment of the present invention;

[0021] Figure 4 A schematic block diagram of the structure of a locomotive equipment maintenance system based on a handheld voiceprint collection terminal provided by an embodiment of the present invention;

[0022] Figure 5A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention.

[0023] Icons: 30-handheld voiceprint collection terminal; 301-microphone array; 302-sensor; 303-communication module; 304-display screen; 305-battery; 306-handle. DETAILED DESCRIPTION

[0024] 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.

[0025] The specific implementation modes of the present invention are described in detail below in conjunction with the accompanying drawings.

[0026] In order to solve the technical problems in the aforementioned background technology, Figure 1 This is a flow chart of a locomotive equipment maintenance method based on a handheld voiceprint collection terminal provided in an embodiment of the present disclosure. The locomotive equipment maintenance method based on a handheld voiceprint collection terminal is introduced in detail below.

[0027] Step S201, obtaining the current equipment voiceprint and environmental parameters collected by the handheld voiceprint collection terminal at a preset locomotive equipment maintenance position;

[0028] Step S202, inputting the current device voiceprint into a voiceprint abnormal type recognition model to obtain a current abnormal voiceprint recognition result corresponding to the current device voiceprint;

[0029] Step S203, generating a current locomotive equipment abnormality description according to the current abnormal voiceprint recognition result and the environmental parameters;

[0030] Step S204, obtaining an inferred locomotive component abnormality according to the current locomotive equipment abnormality description;

[0031] Step S205, obtaining a candidate locomotive equipment maintenance plan corresponding to the inferred locomotive component abnormality from a preset maintenance database as a current locomotive equipment maintenance plan.

[0032] In an embodiment of the present invention, illustratively, the server receives voiceprint data and environmental parameters collected from the engine room of a locomotive by a handheld voiceprint collection terminal. For example, in the engine room of a certain type of electric locomotive, maintenance personnel use a handheld voiceprint collection terminal to record the sound of the engine running, and at the same time collect the temperature, humidity and other environmental parameters in the room. For another example, six data collection sites can be placed in the machine room, and the handheld voiceprint collection terminal is directed to the detected components to detect the acoustic information of the traction motor cooling fan, cooling tower fan, compressor, etc. These data are transmitted to the server in real time via a wireless network. The server inputs the received engine operation voiceprint data into the trained voiceprint abnormal type recognition model. The model analyzes the voiceprint features through a deep learning algorithm and identifies abnormal sound patterns. For example, the model detects irregular metal friction sounds in the voiceprint, which usually indicates wear or misalignment of mechanical parts. Combining the voiceprint abnormality recognition results (metal friction sounds) and environmental parameters (such as high temperature and high humidity in the engine room), the server generates a locomotive equipment abnormality description report. The report pointed out that in a high temperature and high humidity environment, the locomotive engine parts may have problems with wear or misalignment, resulting in abnormal metal friction sounds. Based on the abnormal description report, the server further infers the specific abnormal conditions of the locomotive parts. For example, based on the metal friction sound and high temperature environment, the server can infer that a certain bearing or gear of the engine has worn out, or that the lubricating oil performance has decreased due to high temperature, exacerbating the friction of mechanical parts. Based on the inferred abnormal conditions of the locomotive parts, the server searches for the corresponding maintenance plan in the preset maintenance database. For example, for the problem of bearing wear, the database can provide several different maintenance plans such as replacing bearings and adding high-performance lubricants. The server selects the most appropriate plan, such as replacing worn bearings immediately, and recommends that locomotive maintenance personnel take corresponding actions. In this way, through the intelligent analysis of the handheld voiceprint collection terminal and the server, the fault of the locomotive equipment can be quickly and accurately located, and an effective maintenance plan can be provided.

[0033] In the embodiment of the present invention, the voiceprint abnormality type recognition model is obtained by the following method, including:

[0034] Segmenting the first device voiceprint instance, and performing a feature extraction operation on each segmented voiceprint segment to obtain an original voiceprint feature data set, wherein the original voiceprint feature data set contains an original voiceprint feature corresponding to each voiceprint segment, and the first device voiceprint instance is a device voiceprint instance of a pre-configured category but not configured with an abnormal type;

[0035] Reorganizing and arranging the original voiceprint features in the original voiceprint feature data set to obtain a first voiceprint feature data set and a second voiceprint feature data set, wherein the first voiceprint feature in the first voiceprint feature data set and the second voiceprint feature in the second voiceprint feature data set correspond to different reorganization and arrangement strategies;

[0036] Pre-training a voiceprint abnormality type recognition model according to the first voiceprint feature data set and the second voiceprint feature data set, wherein the voiceprint abnormality type recognition model is used to identify abnormal situations in voiceprints;

[0037] The pre-trained voiceprint abnormality type recognition model is tuned according to the second device voiceprint instance, where the second device voiceprint instance is a device voiceprint instance of a pre-configured category.

[0038] In the embodiment of the present invention, the server first obtains a large number of first device voiceprint instances, which are pre-collected device sound data that are not annotated with abnormal types. Then, the server uses a professional voiceprint analysis tool to accurately segment these voiceprint instances and decompose each voiceprint instance into multiple shorter voiceprint segments. Subsequently, feature extraction is performed on each voiceprint segment to extract key features that can represent the sound characteristics of the segment, such as frequency, amplitude, timbre, etc. The purpose of this step is to construct an original voiceprint feature data set containing various voiceprint segments and their corresponding features. For example, the server can segment the engine running soundprint of an electric locomotive to obtain multiple voiceprint segments with a duration of several seconds, and extract features such as frequency fluctuations and amplitude changes from these segments. After obtaining a large number of original voiceprint features, the server will reorganize and arrange these features to form two different voiceprint feature data sets: a first voiceprint feature data set and a second voiceprint feature data set. The two data sets contain the same voiceprint features, but their organization and order are different to reflect different reorganization and arrangement strategies. The purpose of this is to increase the generalization ability of the model so that it can learn voiceprint features from different perspectives. For example, the first voiceprint feature dataset can organize features in chronological order, while the second voiceprint feature dataset can randomly shuffle the features or rearrange them in a specific pattern. Using these two reorganized voiceprint feature datasets, the server begins to pre-train a voiceprint abnormality type recognition model. This model can be a deep learning network that recognizes abnormalities in sound by learning a large number of voiceprint features. By training on different voiceprint feature datasets, the model can more comprehensively understand the various change patterns of sound and learn to distinguish between normal and abnormal sounds. For example, the model can learn the sound features of the engine when it is running normally on the first voiceprint feature dataset, and learn more types of sound change patterns, including some sound features in abnormal situations, on the second voiceprint feature dataset. After completing the pre-training of the model, the server will use the second device voiceprint instance to further tune the model. These second device voiceprint instances are device sound data that have been labeled with abnormal types, which can help the model better learn and recognize various abnormal sounds. By training on these annotated data, the model can more accurately identify anomalies in voiceprints and give corresponding recognition results. For example, the server can use some locomotive engine voiceprints that are known to have bearing wear or gear failures as second device voiceprint instances to tune the model. In this way, when the model encounters similar sound features again, it can more accurately identify these anomalies.

[0039] In the embodiment of the present invention, the reorganization and arrangement of the original voiceprint features in the original voiceprint feature data set to obtain the first voiceprint feature data set and the second voiceprint feature data set can be implemented through the following examples.

[0040] Adjusting the feature arrangement order of the original voiceprint features in the original voiceprint feature data set to obtain a first original voiceprint feature data set and a second original voiceprint feature data set, wherein the arrangement order of the original voiceprint features in the first original voiceprint feature data set and the second original voiceprint feature data set are different;

[0041] Reorganizing the first original voiceprint feature data set sequence to obtain a first voiceprint feature vector, and reorganizing the second original voiceprint feature data set sequence to obtain a second voiceprint feature vector;

[0042] Performing feature integration on the original voiceprint features in the first voiceprint feature vector, and obtaining the first voiceprint feature data set according to the feature integration result;

[0043] The original voiceprint features in the second voiceprint feature vector are integrated, and the second voiceprint feature data set is obtained according to the feature integration result.

[0044] In the embodiment of the present invention, exemplarily, the server first obtains the original voiceprint feature data set, which contains various voiceprint features extracted from the device sound. In order to increase the generalization ability of the model and learn different feature combinations, the server will adjust the feature arrangement order of these original voiceprint features. Specifically, the server will randomly or according to certain rules to rearrange these features, thereby generating two new data sets: a first original voiceprint feature data set and a second original voiceprint feature data set. The feature arrangement order in these two data sets is different to provide different data perspectives for model learning. For example, assuming that the original voiceprint feature data set contains three features of frequency, amplitude and duration, the original arrangement may be [frequency, amplitude, duration]. After adjustment, the arrangement of the first original voiceprint feature data set may be [amplitude, duration, frequency], and the arrangement of the second original voiceprint feature data set may be [duration, frequency, amplitude]. After adjusting the feature arrangement order, the server will reorganize the voiceprint features according to the new sequence to form voiceprint feature vectors. These vectors will be used as input for subsequent model training. For the first original voiceprint feature data set and the second original voiceprint feature data set, the server generates a first voiceprint feature vector and a second voiceprint feature vector respectively. Continuing with the above example, the first voiceprint feature vector can be a vector arranged in the order of [amplitude, duration, frequency], while the second voiceprint feature vector can be a vector arranged in the order of [duration, frequency, amplitude]. In order to further improve the performance of the model, the server integrates the original voiceprint features in the voiceprint feature vector. This can be achieved by combining, transforming or extracting higher-level features. The purpose of feature integration is to better capture the inherent laws and patterns in the voiceprint data. For example, for the two features of frequency and amplitude, the server can calculate their ratio or difference to form a new integrated feature. This new feature may be more helpful for the model to identify abnormalities in the sound. After completing the feature integration, the server generates two new voiceprint feature data sets based on the integration results: the first voiceprint feature data set and the second voiceprint feature data set. These two data sets will contain rearranged and integrated voiceprint features, providing richer and more useful information for subsequent model training. Ultimately, these two datasets will be used to train and tune the voiceprint anomaly type recognition model, enabling it to more accurately identify and classify anomalies in device sounds.

[0045] In the embodiment of the present invention, the step of integrating the original voiceprint features in the first voiceprint feature vector and obtaining the first voiceprint feature data set according to the feature integration result may be implemented through the following examples.

[0046] Selecting a plurality of continuous original voiceprint features in the first voiceprint feature vector by a continuous feature selector;

[0047] Performing feature integration on the multiple original voiceprint features to obtain a first integrated voiceprint feature;

[0048] Performing feature conversion on multiple groups of the first integrated voiceprint features to obtain the first voiceprint feature data set, wherein the multiple groups of the first integrated voiceprint features are obtained by shifting the continuous feature selector;

[0049] The step of integrating the original voiceprint features in the second voiceprint feature vector and obtaining the second voiceprint feature data set according to the feature integration result may be implemented through the following example.

[0050] Selecting a plurality of continuous original voiceprint features in the second voiceprint feature vector by a continuous feature selector;

[0051] Performing feature integration on the multiple original voiceprint features to obtain a second integrated voiceprint feature;

[0052] Feature conversion is performed on multiple groups of the second integrated voiceprint features to obtain the second voiceprint feature data set, and the multiple groups of the second integrated voiceprint features are obtained by shifting the continuous feature selector.

[0053] In the embodiment of the present invention, exemplarily, the server first uses a continuous feature selector to select a continuous original voiceprint feature in the first voiceprint feature vector. For example, assuming that the first voiceprint feature vector contains multiple features such as frequency, amplitude, waveform complexity, etc., the continuous feature selector can first select two continuous features of frequency and amplitude. The server then integrates the selected continuous original voiceprint features. Taking frequency and amplitude as an example, the server can calculate the weighted average, ratio or other statistics of these two features to form a new integrated feature, namely the first integrated voiceprint feature. This new feature may more comprehensively reflect certain characteristics of the sound signal. In order to obtain more diverse feature combinations, the server will select different combinations of original voiceprint features by shifting the continuous feature selector and repeat the above integration process. In this way, the server will obtain multiple groups of first integrated voiceprint features. Subsequently, the server will perform further feature conversion on these integrated features, such as standardization, normalization or application of some mathematical transformation (such as Fourier transform) to form the final first voiceprint feature data set. Similar to processing the first voiceprint feature vector, the server will also use a continuous feature selector to select a continuous original voiceprint feature in the second voiceprint feature vector. Since the feature arrangement order of the second voiceprint feature vector is different from that of the first voiceprint feature vector, the selected feature combination will also be different. The server integrates the continuous original voiceprint features selected from the second voiceprint feature vector to form a second integrated voiceprint feature. This process is similar to the processing of the first voiceprint feature vector, but due to the different input features, the integrated features obtained will also be different. Similarly, the server will obtain different combinations of original voiceprint features by shifting the continuous feature selector and generate multiple groups of second integrated voiceprint features. Afterwards, these integrated features are feature transformed to construct a second voiceprint feature data set. This data set will be used together with the first voiceprint feature data set to train and optimize the voiceprint anomaly type recognition model.

[0054] In the embodiment of the present invention, the feature integration of the multiple original voiceprint features to obtain the first integrated voiceprint feature can be implemented through the following examples.

[0055] Performing feature concatenation on the multiple original voiceprint features to obtain the first integrated voiceprint feature, or performing feature integration on the multiple original voiceprint features to obtain the first integrated voiceprint feature;

[0056] The step of integrating the plurality of original voiceprint features to obtain a second integrated voiceprint feature includes:

[0057] The plurality of original voiceprint features are concatenated to obtain the second integrated voiceprint feature, or the plurality of original voiceprint features are integrated to obtain the second integrated voiceprint feature.

[0058] In an embodiment of the present invention, illustratively, the server first selects multiple original voiceprint features in the first voiceprint feature vector, such as frequency, amplitude and waveform complexity. These features reflect different aspects of the locomotive sound, such as the engine speed, vibration intensity and complexity of the sound pattern. Through feature concatenation, the server combines these feature values ​​into a longer feature sequence in a specific order. For example, assuming that the value of the frequency feature is 500Hz, the value of the amplitude feature is 0.8, and the value of the waveform complexity is 0.6, the server can concatenate these three feature values ​​into a new feature vector: [500, 0.8, 0.6]. This new feature vector is the first integrated voiceprint feature, which integrates the information of multiple original voiceprint features and provides a more comprehensive data basis for subsequent voiceprint analysis. In addition to concatenation, the server can also choose to integrate multiple original voiceprint features. For example, the server can calculate the product or ratio of frequency and amplitude to form a new feature. This integrated feature may more directly reflect a specific state of the locomotive, such as the load condition or vibration mode of the engine. Assuming that the value of the frequency feature is 500Hz and the value of the amplitude feature is 0.8, the server can calculate their product as a new integrated feature: 500*0.8=400. This new integrated feature value 400 is part of the first integrated voiceprint feature, which represents the characteristics of the locomotive sound in a new way. For the original voiceprint features in the second voiceprint feature vector, the server will perform similar operations, but because the feature arrangement order of the second voiceprint feature vector is different from that of the first voiceprint feature vector, the integrated second integrated voiceprint feature will also be different. In the second voiceprint feature vector, the server may first select the three features of amplitude, waveform complexity and frequency (note that the order is different from that in the first voiceprint feature vector). Through concatenation, the server can obtain a new feature vector, such as [0.8, 0.6, 500], as the second integrated voiceprint feature. Similarly, in the second voiceprint feature vector, the server can also integrate amplitude, waveform complexity and frequency. For example, the ratio of the sum of amplitude and waveform complexity to frequency can be calculated to obtain a new integrated feature. This new integrated feature will be part of the second integrated voiceprint feature, providing another perspective for voiceprint analysis. Through these feature integration methods, the server can more comprehensively capture and analyze multiple characteristics in locomotive sounds, thereby improving the accuracy and efficiency of equipment status monitoring and fault diagnosis.

[0059] In the embodiment of the present invention, the reorganization according to the first original voiceprint feature data set sequence to obtain the first voiceprint feature vector, and the reorganization according to the second original voiceprint feature data set sequence to obtain the second voiceprint feature vector, can be implemented through the following examples.

[0060] Determining a vector dimension according to a segmentation strategy of the voiceprint instance of the first device;

[0061] According to the vector dimension, the original voiceprint features in the first original voiceprint feature data set are sequenced to obtain the first voiceprint feature vector;

[0062] According to the vector dimension, the original voiceprint features in the second original voiceprint feature data set are sequenced to obtain the second voiceprint feature vector.

[0063] In the embodiment of the present invention, illustratively, the server first determines the dimension of the voiceprint feature vector according to the segmentation strategy of the voiceprint instance of the first device. For example, if the segmentation strategy is to evenly divide each voiceprint instance into 10 time segments, and extract 5 key voiceprint features (such as frequency, amplitude, waveform complexity, sound intensity and sound quality) for each time segment, the dimension of the voiceprint feature vector will be determined to be 50 (10 time segments × 5 voiceprint features / segment). After determining the vector dimension, the server will start processing the first original voiceprint feature data set. This data set contains unprocessed voiceprint data obtained from a handheld voiceprint acquisition terminal. The server will reorder the original voiceprint features in the data set according to the segmentation strategy. Specifically, it will arrange the 5 key voiceprint features extracted in each time segment in a specific order (such as frequency, amplitude, waveform complexity, sound intensity, sound quality), and fill these feature values ​​into a 50-dimensional vector in turn. In this way, the server obtains a structured first voiceprint feature vector, which comprehensively describes the voiceprint features of the locomotive in a form that is convenient for subsequent analysis and processing. The server performs similar processing on the second original voiceprint feature data set. However, since the voiceprint features in the second data set may be different from those in the first data set (for example, the feature extraction method, order or weight may have changed), the second voiceprint feature vector obtained will also be different from the first vector. In the process of reorganizing the second voiceprint feature vector, the server will also follow the segmentation strategy and determine the corresponding vector dimension. Then, it will fill the original voiceprint features in the second data set into the vector according to the new feature order or extraction method. In this way, the server obtains two different but interrelated voiceprint feature vectors, which describe the running status of the locomotive from different perspectives. Through this reorganization method, the server can effectively convert the original and messy voiceprint data into structured and easy-to-analyze voiceprint feature vectors. These vectors not only contain rich locomotive operation information, but also provide strong data support for subsequent voiceprint recognition, fault diagnosis and other tasks.

[0064] In the embodiment of the present invention, after the original voiceprint features in the original voiceprint feature data set are reorganized and arranged to obtain the first voiceprint feature data set and the second voiceprint feature data set, the embodiment of the present invention also provides the following implementation manner.

[0065] According to the first voiceprint feature data set, cyclically performing reorganization and arrangement at least once to obtain a third voiceprint feature data set;

[0066] According to the second voiceprint feature data set, cyclically performing reorganization and arrangement at least once to obtain a fourth voiceprint feature data set;

[0067] A voiceprint abnormality type recognition model is pre-trained according to the third voiceprint feature data set and the fourth voiceprint feature data set.

[0068] In an embodiment of the present invention, exemplarily, after the server completes the initial reorganization and arrangement of the original voiceprint feature data set to obtain the first voiceprint feature data set, it will further perform a cyclic reorganization and arrangement on the data set. The purpose of this cyclic reorganization and arrangement is to increase the diversity and complexity of the data set, so as to make the trained model more robust. For example, the server can adopt the following strategies for cyclic reorganization and arrangement: Random shuffling: Randomly shuffle the feature order in the first voiceprint feature data set to destroy the temporal dependency in the original data and force the model to learn a more robust feature representation. Feature addition and deletion: Randomly delete or add certain features to simulate data missing or redundancy that may occur in actual applications. Noise injection: Add a certain amount of noise to the data set to simulate interference factors that may occur in actual applications. Through the above strategies, the server obtains a third voiceprint feature data set, which adds more changes and challenges on the basis of the first voiceprint feature data set. Similar to the generation of the third voiceprint feature data set, the server will also perform a cyclic reorganization and arrangement on the second voiceprint feature data set to obtain a fourth voiceprint feature data set. In this process, the server can adopt different reorganization strategies or adjust the parameters of existing strategies to ensure that the two data sets have certain differences while maintaining relevance. After obtaining the third and fourth voiceprint feature data sets, the server will use these two data sets to pre-train the voiceprint abnormal type recognition model. The purpose of this pre-training is to enable the model to accurately identify the abnormal type when facing diversified and complex voiceprint data. For example, the server can use a deep learning model (such as a convolutional neural network or a recurrent neural network) as the infrastructure of the voiceprint abnormal type recognition model. During the training process, the server will continuously input samples from the third and fourth voiceprint feature data sets into the model, and adjust the parameters of the model based on the difference between the model's output and the true label. In this way, the server can train a recognition model that is highly sensitive and accurate to voiceprint abnormal types, providing strong support for subsequent practical applications.

[0069] In the embodiment of the present invention, the pre-training of the voiceprint abnormality type recognition model according to the first voiceprint feature data set and the second voiceprint feature data set can be implemented through the following examples.

[0070] Loading the first voiceprint feature data set into the dynamic network branch of the voiceprint abnormal type recognition model to obtain a first abnormal voiceprint recognition result;

[0071] Loading the second voiceprint feature data set into the static network branch of the voiceprint abnormal type recognition model to obtain a second abnormal voiceprint recognition result;

[0072] Training the dynamic network branch according to the first abnormal voiceprint recognition result and the second abnormal voiceprint recognition result;

[0073] According to the network parameters of the dynamic network branch that have completed training, the network parameters of the static network branch are adjusted.

[0074] In an embodiment of the present invention, exemplarily, the server first loads the first voiceprint feature data set into the dynamic network branch of the voiceprint abnormal type recognition model. This dynamic network branch is designed to capture dynamic changes in voiceprint features to identify possible anomalies. For example, the server can input each sample in the first voiceprint feature data set into the dynamic network branch in turn, and obtain the first abnormal voiceprint recognition result of each sample through the forward propagation of the network. These results may indicate whether there is an abnormal voiceprint in the sample, as well as information such as the specific type of the abnormality. Similar to the dynamic network branch, the server then loads the second voiceprint feature data set into the static network branch of the voiceprint abnormal type recognition model. This static network branch focuses more on capturing static information in the voiceprint features to assist the dynamic network branch in more accurate abnormal identification. The server will also input each sample in the second voiceprint feature data set into the static network branch in turn, and obtain the second abnormal voiceprint recognition result of each sample through the forward propagation of the network. After obtaining the first abnormal voiceprint recognition result and the second abnormal voiceprint recognition result, the server will use these two results to train the dynamic network branch. Specifically, the server compares the difference between the first abnormal voiceprint recognition result and the real abnormal label, and then updates the network parameters of the dynamic network branch through the back propagation algorithm to minimize the difference. For example, if a sample is identified as having an abnormal voiceprint, but in fact there is no abnormality, the server will adjust the parameters of the dynamic network branch to reduce the misjudgment of the sample. When the dynamic network branch is trained, the server will adjust the network parameters of the static network branch according to its trained network parameters. This is because although the dynamic network branch and the static network branch focus on different information, their recognition goals are the same, that is, to accurately identify the abnormalities in the voiceprint. The server can use the transfer learning method to transfer the knowledge learned in the dynamic network branch to the static network branch, thereby improving the recognition performance of the static network branch. For example, the server can copy the parameters of certain layers (such as convolutional layers or fully connected layers) in the dynamic network branch to the corresponding layers of the static network branch as the initial parameters of the static network branch, and then conduct further training and optimization on this basis.

[0075] In the embodiment of the present invention, the training of the dynamic network branch according to the first abnormal voiceprint recognition result and the second abnormal voiceprint recognition result can be implemented through the following examples.

[0076] Determine a matching degree error between the first abnormal voiceprint recognition result and the second abnormal voiceprint recognition result;

[0077] According to the matching degree and matching degree error, the network parameters of the dynamic network branch are adjusted by error feedback.

[0078] In an embodiment of the present invention, illustratively, the server first compares the first abnormal voiceprint recognition result with the second abnormal voiceprint recognition result to determine the degree of match between them. The calculation of this degree of match may be based on a variety of indicators, such as accuracy, recall rate, F1 score, etc. For example, the server can calculate the recognition consistency of the same abnormal type in the two recognition results. If the first abnormal voiceprint recognition result considers that a sample belongs to abnormal type A, while the second abnormal voiceprint recognition result considers that the sample belongs to abnormal type B, then this indicates that there is a mismatch between the two recognition results. The server will quantify this mismatch as a specific matching error. This error value reflects the degree of divergence between the dynamic network branch and the static network branch in abnormal voiceprint recognition. After determining the matching error, the server will use this error value to adjust the network parameters of the dynamic network branch through an error backpropagation algorithm. Specifically, the server will calculate the error gradient, that is, the partial derivative of the matching error for each parameter in the dynamic network branch. This gradient indicates how each parameter should be adjusted in order to reduce the matching error. Then, the server will update the parameters of the dynamic network branch according to the calculated gradient. This process is usually achieved through optimization algorithms (such as gradient descent, Adam, etc.). By continuously iterating this process, the server can gradually reduce the matching error, thereby improving the accuracy of the dynamic network branch in abnormal voiceprint recognition. For example, if the dynamic network branch mistakenly identifies a normal voiceprint as an abnormal voiceprint, then by adjusting the network parameters, the server can enable the dynamic network branch to make a more accurate judgment the next time it encounters a similar situation.

[0079] In the embodiment of the present invention, adjusting the network parameters of the static network branch according to the trained network parameters of the dynamic network branch can be implemented through the following examples.

[0080] According to the network parameters of the dynamic network branch that have completed training, the network parameters of the static network branch are adjusted through model integration.

[0081] In the embodiment of the present invention, for example, model integration is an effective means to improve model performance, which can improve the overall prediction accuracy by combining the prediction results of multiple models. In this scenario, the server will use the parameters of the trained dynamic network branch to adjust the model integration of the static network branch. Specifically, the server can take the following steps: Parameter sharing: The server will first analyze the trained parameters in the dynamic network branch, especially those layers or modules that are critical for abnormal voiceprint recognition. Then, it will share these important parameters with the static network branch as the initial parameters of the static network branch or to replace the corresponding parameters in the static network branch. Feature fusion: In addition to directly sharing parameters, the server can also consider fusing the features of the dynamic network branch and the static network branch. For example, the outputs of the middle layers of the two branches can be spliced, weighted summed, or other forms of fusion to make full use of the different feature information extracted by the two branches. This feature fusion can enhance the recognition ability of the static network branch for complex voiceprint features. Joint training: After parameter sharing and feature fusion, the server can further perform joint training on the static network branch. This means that during the training process, the static network branch will not only receive error feedback information from itself, but also receive auxiliary information from the dynamic network branch. Through this joint training method, the static network branch can better learn useful knowledge from the dynamic network branch, thereby improving its own recognition performance. For example, in the voiceprint recognition of locomotive equipment, the dynamic network branch may be better at capturing the dynamic change characteristics of the equipment during operation, while the static network branch pays more attention to the inherent characteristics of the equipment. Through model integration adjustment, the server can combine the advantages of both, so that the static network branch can perform well in the face of various complex voiceprint recognition tasks.

[0082] In the embodiment of the present invention, the tuning of the pre-trained voiceprint abnormality type recognition model according to the voiceprint instance of the second device can be implemented through the following examples.

[0083] The network parameters of the static network branch in the voiceprint anomaly type recognition model are tuned according to the voiceprint instance of the second device.

[0084] In an embodiment of the present invention, illustratively, first, the server collects voiceprint data from the actual running locomotive equipment, and these data contain the voiceprint information of the equipment in various states, such as normal operation, abnormal operation, etc. These data are called second device voiceprint instances, and they are crucial for model tuning. Before tuning, the server preprocesses the collected second device voiceprint instances. This includes operations such as data cleaning, standardization, and normalization to ensure the quality and consistency of the data. The preprocessed data will be more suitable for model tuning. Next, the server uses the second device voiceprint instance to tune the static network branch in the voiceprint abnormal type recognition model. The static network branch is mainly responsible for capturing the static information in the voiceprint feature. During the tuning process, the server inputs the second device voiceprint instance into the model and obtains the prediction result of the model through forward propagation. Then, the server compares the difference between the prediction result and the true label and calculates the error. Based on this error, the server uses an optimization algorithm (such as gradient descent) to adjust the network parameters of the static network branch. This adjustment process is achieved through error backpropagation, that is, the gradient of the parameter is calculated based on the error, and the parameter value is updated in the direction of the gradient. For example, if a certain parameter plays a negative role in the model prediction (causing the predicted result to deviate greatly from the actual result), then during the tuning process, the value of this parameter will be adjusted accordingly to reduce its negative impact on the model prediction. The server will repeatedly perform the above tuning process until the performance of the model reaches a satisfactory level or reaches a preset number of iterations. After each tuning, the server will use the validation data set to evaluate the performance of the model to ensure that the tuning is effective. In this way, the server can use the second device voiceprint instance to fine-tune the static network branch in the voiceprint anomaly type recognition model, thereby improving the accuracy and robustness of the model in actual application scenarios.

[0085] In the embodiment of the present invention, the tuning of the network parameters of the static network branch in the voiceprint anomaly type recognition model according to the voiceprint instance of the second device can be implemented through the following examples.

[0086] Loading the second device voiceprint instance to the static network branch of the voiceprint abnormal type recognition model to obtain a sample abnormal voiceprint recognition result;

[0087] According to the sample abnormal voiceprint recognition result and the sample abnormal type configuration corresponding to the second device voiceprint instance, the network parameters of the static network branch are adjusted through error feedback.

[0088] In an embodiment of the present invention, exemplarily, first, the server will load the second device voiceprint instance into the static network branch of the voiceprint abnormal type recognition model. These voiceprint instances are collected from the actual running locomotive equipment and contain voiceprint data of various abnormal types. For example, the server may load a sample containing a bearing wear abnormal voiceprint into the static network branch. This sample will pass through the forward propagation of the network to obtain a sample abnormal voiceprint recognition result. The static network branch will process the loaded second device voiceprint instance and output a sample abnormal voiceprint recognition result. This result can be a probability distribution, indicating the possibility that the sample belongs to different abnormal types. Continuing with the above example, the static network branch can output a probability distribution indicating that the sample has an 80% probability of bearing wear abnormality, a 10% probability of other types of abnormalities, and a 10% probability of normal voiceprints. Next, the server will obtain the sample abnormal type configuration corresponding to the second device voiceprint instance. This configuration is pre-marked and indicates the true abnormal type of each sample. In the above example, the server will know that the loaded sample is actually a bearing wear abnormality. The server then compares the difference between the sample abnormal voiceprint recognition results and the actual abnormal type. In this example, although the model identified the highest probability of bearing wear abnormality, there is still a 20% error. In order to reduce this error, the server will tune the network parameters of the static network branch through the error backpropagation algorithm. Specifically, the server calculates the error gradient, that is, the partial derivative of the error between the predicted result and the actual result with respect to each network parameter, and updates the network parameters based on this gradient. In this way, the server can gradually adjust the parameters of the static network branch to more accurately identify various types of abnormal voiceprints. This process is repeated until the performance of the model reaches a satisfactory level.

[0089] In the embodiment of the present invention, the inferred locomotive component abnormality according to the current locomotive equipment abnormality description may be implemented through the following examples.

[0090] Extracting target description content of target locomotive component abnormality including current locomotive equipment abnormality description to obtain target description content;

[0091] Obtaining a hypergraph for representing the association relationship between the plurality of reference description contents, and extracting from the hypergraph a local hypergraph corresponding to each of the target description contents;

[0092] There is an association relationship between the entity in the hypergraph and the reference description content, the local hypergraph includes a target entity, and the matching degree between the reference description content corresponding to the target entity and the corresponding target description content exceeds a matching degree threshold;

[0093] Extracting topological features from each of the local hypergraphs to obtain topological features of each of the local hypergraphs, and fusing features of each of the topological features to obtain fused topological features of the abnormal conditions of the target locomotive components;

[0094] Based on the fused topological features, the abnormality of the target locomotive component is inferred to obtain an inferred locomotive component abnormality corresponding to the target locomotive component abnormality, and the current locomotive equipment abnormality description of the inferred locomotive component abnormality is consistent with the current locomotive equipment abnormality description of the target locomotive component abnormality.

[0095] In an embodiment of the present invention, exemplarily, the server first extracts the target description content of the target locomotive component abnormality including the current locomotive equipment abnormality description. For example, the abnormal description can be "the locomotive engine sound is abnormal, accompanied by obvious metal friction sound", and the server extracts key description content from it, such as "the engine sound is abnormal" and "the metal friction sound". Next, the server obtains a hypergraph for characterizing the association relationship between multiple reference description contents. This hypergraph is constructed based on historical data and expert knowledge, in which entities represent different abnormal descriptions or fault types, and hyperedges represent the relationship between these descriptions or types. Then, the server extracts the local hypergraph corresponding to each target description content from the hypergraph. For example, for "the engine sound is abnormal", the server finds all entities and relationships associated with it to form a local hypergraph. Similarly, for "the metal friction sound", a corresponding local hypergraph is also formed. The reference description content corresponding to the entities in these local hypergraphs needs to match the target description content to a degree that exceeds a preset matching degree threshold. For each local hypergraph, the server extracts topological features, such as the degree of the node, the connection of the hyperedge, etc. These topological features can reflect the complex relationship between the abnormal descriptions. After extracting the topological features, the server will fuse the topological features of all local hypergraphs to obtain a fused topological feature. This fused feature comprehensively considers the association between all target description contents in the hypergraph, so it can more comprehensively characterize the abnormal conditions of the target locomotive parts. Finally, based on the fused topological features, the server will infer the abnormal conditions of the target locomotive parts. By comparing the fused features with the features in historical data or preset models, the server can determine the specific faults or abnormal conditions that may exist in the target locomotive parts. For example, based on the fused topological features, the server can infer that the actual fault corresponding to the description of "abnormal locomotive engine sound, accompanied by obvious metal friction sound" is "engine bearing wear". This inference result will be consistent with the original description of the abnormal condition of the target locomotive parts, thereby helping maintenance personnel to accurately and quickly locate and solve the problem.

[0096] In the embodiment of the present invention, the target description content is extracted from the target locomotive component abnormality including the current locomotive equipment abnormality description to obtain the target description content, which can be implemented through the following examples.

[0097] Acquire the abnormal condition of the locomotive component to be analyzed, and identify the abnormal description of the current locomotive equipment for the abnormal condition of the locomotive component to be analyzed, so as to obtain the abnormal description information of the current locomotive equipment for the abnormal condition of the locomotive component to be analyzed;

[0098] When the current locomotive equipment abnormality description information indicates that the locomotive component abnormality to be analyzed includes the current locomotive equipment abnormality description, determining the locomotive component abnormality to be analyzed as the target locomotive component abnormality;

[0099] Performing semantic analysis on the abnormal condition of the target locomotive component to obtain multiple key descriptions of the abnormal condition of the target locomotive component;

[0100] For each of the key descriptions, when the key description includes the current locomotive equipment abnormality description, the key description is determined as the target description content.

[0101] In the embodiment of the present invention, exemplarily, first, the server will obtain a record of abnormal conditions of a locomotive component to be analyzed from the database. For example, this record can be a fault report about "abnormal noise when the locomotive engine is running". Next, the server will automatically identify this fault report and extract the abnormal description information of the current locomotive equipment. In this example, the server can identify "abnormal noise" as the abnormal description of the current locomotive equipment. Once the server identifies the abnormal description information of the current locomotive equipment, it will check whether this description is indeed included in the abnormal condition of the locomotive component to be analyzed. If it is included, then the abnormal condition of the locomotive component to be analyzed is determined as the target locomotive component abnormal condition. In the example of the embodiment of the present invention, the fault report of "abnormal noise when the locomotive engine is running" is determined as the target locomotive component abnormal condition. Then, the server will perform semantic analysis on the abnormal condition of the target locomotive component and extract multiple key descriptions. In this example, the key descriptions can include "engine running", "abnormal noise", etc. Finally, for each key description, the server will check whether it contains the abnormal description of the current locomotive equipment. If it is included, the key description is determined as the target description content. In the example of the embodiment of the present invention, the key description "abnormal noise" is determined as the target description content because it directly corresponds to the abnormal description of the current locomotive equipment. Through this series of steps, the server can accurately extract the target description content related to the abnormal description of the current locomotive equipment from the complex abnormal conditions of the locomotive parts, providing an important basis for subsequent fault diagnosis and repair work.

[0102] In the embodiment of the present invention, after the current locomotive equipment abnormality description is identified for the abnormality of the locomotive component to be analyzed and the current locomotive equipment abnormality description information of the abnormality of the locomotive component to be analyzed is obtained, the following implementation manner is also provided.

[0103] When the current locomotive equipment abnormality description information indicates that the abnormality of the locomotive component to be analyzed does not include the current locomotive equipment abnormality description, respectively extracting the reference hypergraphs corresponding to each of the key descriptions from the hypergraph, performing topological feature extraction on each of the reference hypergraphs to obtain reference topological features of each of the reference hypergraphs, and performing feature fusion on each of the reference topological features to obtain reference fused topological features of the abnormality of the locomotive component to be analyzed;

[0104] Based on the reference fusion topological features, abnormality inference is performed on the abnormality of the locomotive component to be analyzed to obtain a target inferred locomotive component abnormality corresponding to the abnormality of the locomotive component to be analyzed;

[0105] The target inferred locomotive component abnormality includes the current locomotive equipment abnormality description, and the target inferred locomotive component abnormality includes the locomotive component abnormality to be analyzed.

[0106] In the embodiment of the present invention, for example, it is assumed that the server is processing an abnormality of a locomotive component to be analyzed, and the abnormality is about a fault in the locomotive electrical system. The preliminary report does not directly mention the specific abnormal description of the current locomotive equipment, such as "voltage fluctuation" or "current instability". The server will first identify the abnormal description of the current locomotive equipment for the abnormality of the locomotive component to be analyzed, but in this scenario, the identification result shows that the abnormality of the locomotive component to be analyzed does not directly contain the abnormal description information of the current locomotive equipment. Since the abnormality to be analyzed does not contain a clear description of the equipment abnormality, the server will extract the reference hypergraph corresponding to each key description related to the abnormality from the previously constructed hypergraph. For example, the key description may include "electrical system failure", "power module abnormality", etc., and the server will extract the corresponding reference hypergraph from the hypergraph for these descriptions. For each extracted reference hypergraph, the server will extract topological features, which may include node connectivity, clustering coefficient, etc. Subsequently, the server will fuse the topological features of all reference hypergraphs to obtain a reference fusion topological feature, which comprehensively reflects the topological structure characteristics of the abnormality of the locomotive component to be analyzed. Based on the obtained baseline fusion topology features, the server will infer the abnormal conditions of the locomotive components to be analyzed. By comparing historical data and models, the server can infer that this abnormal condition is actually related to "voltage instability", although it is not explicitly mentioned in the original report. In the end, the server will obtain a target inferred locomotive component abnormality. This inference result contains the abnormal description of the current locomotive equipment, such as "voltage instability", and this inference result also covers the original abnormal condition of the locomotive component to be analyzed, that is, electrical system failure. In this way, even if the specific abnormality is not clearly stated in the original report, the server can identify potential problems through analysis and inference.

[0107] In the embodiment of the present invention, extracting the local hypergraphs corresponding to each target description content from the hypergraph can be implemented through the following examples.

[0108] Comparing the target description content with each of the reference description contents in the hypergraph respectively to obtain a comparison result, wherein the comparison result is used to indicate whether the hypergraph contains reference description contents that are consistent with the target description content;

[0109] Based on the comparison result, selecting a target reference description content corresponding to the target description content from each of the reference description contents in the hypergraph;

[0110] Determine the entity corresponding to the target reference description content in the hypergraph as the target entity corresponding to the target description content, and determine the entity in the hypergraph that has a connection relationship with the target entity as the reference entity corresponding to the target description content;

[0111] The area in the hypergraph formed by the target entity and the reference entity is determined as a local hypergraph corresponding to the target description content.

[0112] In an embodiment of the present invention, illustratively, it is assumed that the server is processing a situation regarding a failure of a locomotive brake system, and the target description content is "abnormal wear of brake pads". The server has constructed a hypergraph containing a variety of locomotive fault and abnormal descriptions. The server first compares the target description content of "abnormal wear of brake pads" with each reference description content in the hypergraph. This comparison is to find out whether there is a reference description content that is consistent with the target description content in the hypergraph. Based on the comparison result, the server selects the target reference description content that is consistent with "abnormal wear of brake pads" from the reference description content of the hypergraph. For example, there may be a reference description content in the hypergraph that is "excessive wear of brake pads leads to reduced braking performance", which is highly relevant to the target description content. The server then determines the entity corresponding to the target reference description content in the hypergraph as the target entity. In this example, the target entity may represent "abnormal wear of brake pads" or a related fault type. Then, the server identifies other entities that have a connection relationship with this target entity in the hypergraph, and these entities are determined as reference entities. The reference entity may represent other faults, components or influencing factors related to abnormal wear of brake pads. Finally, the server extracts the target entity and the reference entity and the connection relationship between them from the hypergraph and constructs a new local hypergraph. This local hypergraph focuses on "brake pad wear abnormality" and its associated faults and factors, providing a targeted view for subsequent fault diagnosis and solutions. In this way, the server can use the structure and association of the hypergraph to quickly locate and analyze specific locomotive faults or abnormal descriptions, thereby improving the efficiency and accuracy of fault diagnosis.

[0113] In the embodiment of the present invention, the selecting, based on the comparison result, the target reference description content corresponding to the target description content from each of the reference description content in the hypergraph can be implemented through the following examples.

[0114] When the comparison result indicates that the hypergraph contains reference description content that is consistent with the target description content, the reference description content that is consistent with the target description content is determined as the target reference description content corresponding to the target description content;

[0115] When the comparison result indicates that the hypergraph does not contain reference description content that is consistent with the target description content, obtaining a matching degree between the target description content and each of the reference description contents;

[0116] The reference description content whose matching degree exceeds the matching degree threshold is determined as the target reference description content corresponding to the target description content.

[0117] In the embodiment of the present invention, illustratively, the server has obtained the target description content, such as "the temperature of the locomotive engine is too high". The server compares the target description content with the reference description content in the hypergraph. The comparison result shows that the hypergraph does contain reference description content that is completely consistent with the target description content "the temperature of the locomotive engine is too high". During the comparison process, the server found that there is a reference description content in the hypergraph that is "the temperature of the locomotive engine is too high, which may cause cooling system failure", which is consistent with the target description content. Therefore, the server directly determines this consistent reference description content as the target reference description content. The server also obtains the target description content, such as "the locomotive brake system responds slowly". After comparison, the server finds that there is no reference description content in the hypergraph that is completely consistent with the target description content. The server then calculates the degree of match between the target description content and each reference description content. The server sets a matching degree threshold, such as 80%, for screening reference description content that is closer to the target description content. After calculating the matching degree, the server finds that there is a reference description content in the hypergraph that is "the reaction time of the locomotive brake is extended", and its matching degree with the target description content is 85%, which exceeds the set threshold. Therefore, although this reference description content is not completely consistent, the server still determines it as the target reference description content because it is highly related to the target description content. Through the above two scenarios, we can see how the server flexibly selects target reference description content related to the target description content from the hypergraph based on the comparison results, whether it is a direct match or a screening through a matching degree threshold.

[0118] In an embodiment of the present invention, the abnormal situation inference is implemented by a noise filtering unit and an abnormal situation inference unit. Based on the fused topological features, the abnormal situation of the target locomotive component is inferred to obtain the inferred locomotive component abnormal situation corresponding to the target locomotive component abnormal situation. The implementation can be performed through the following examples.

[0119] Based on the noise filtering unit, the fused topological feature is subjected to noise filtering to obtain an optimized feature of the abnormal condition of the target locomotive component;

[0120] Based on the abnormality inference unit, the abnormality inference of the target locomotive component is performed according to the optimization feature to obtain the inferred locomotive component abnormality corresponding to the target locomotive component abnormality.

[0121] In an embodiment of the present invention, illustratively, the server first obtains fused topological features, which are extracted and fused from the hypergraph of the abnormal situation of the target locomotive component. Next, the server calls a noise filtering unit, which is specifically used to process the noise data that may exist in the fused topological features. The noise filtering unit carefully analyzes the fused topological features, identifies and removes the noise caused by data acquisition errors, transmission interference or other reasons. After noise filtering, the server obtains optimized features, which are more accurate and can better reflect the nature of the abnormal situation of the target locomotive component. Assume that the fused topological features obtained by the server contain some abnormal data points caused by temporary sensor failure. Through the processing of the noise filtering unit, these abnormal data points are identified and removed. Finally, the server obtains a set of smoother and more accurate optimized features. After obtaining the optimized features, the server calls the abnormal situation inference unit. This unit will make in-depth inferences on the abnormal situation of the target locomotive component based on the optimized features, combined with historical data and machine learning models. Through complex calculations and analysis, the abnormal situation inference unit can identify the specific type, possible causes and potential solutions of the abnormal situation of the target locomotive component. Finally, the server will get an inferred locomotive component abnormality corresponding to the target locomotive component abnormality. Assume that the server is processing a case about a locomotive engine abnormality. After being processed by the noise filtering unit, the server obtains a set of optimized features. The abnormality inference unit makes inferences based on these optimized features, combined with historical data and machine learning models. Finally, the server infers that the engine abnormality may be caused by a sensor failure and gives corresponding solution suggestions. Through the description of the above two scenarios, we can see how the server can accurately analyze and infer the abnormality of the target locomotive component through the two units of noise filtering and abnormality inference.

[0122] In an embodiment of the present invention, the noise filtering unit includes a plurality of noise filtering components. Based on the noise filtering unit, the fused topological features are subjected to noise filtering to obtain optimized features of the abnormal conditions of the target locomotive components. The implementation may be performed through the following examples.

[0123] The fused topological features are subjected to noise filtering in sequence through a plurality of cascaded noise filtering components to obtain optimized features of the abnormal conditions of the target locomotive components.

[0124] In an embodiment of the present invention, illustratively, the server has obtained the fused topological features of the abnormal conditions of the target locomotive components, which may contain noise introduced by various factors (such as sensor errors, environmental interference, etc.). In order to obtain more accurate and reliable feature data, the server needs to filter and optimize these features through a noise filtering unit. The server first initializes the noise filtering unit, which consists of multiple cascaded noise filtering components. Each component is designed and optimized for different types of noise or data anomalies. The fused topological features first pass through the first-level noise filtering component. This component is mainly responsible for identifying and removing high-frequency noise caused by sensor failure or transient interference. For example, a sensor may produce a momentary abnormal reading due to electromagnetic interference, which appears as a sharp peak in the fused topological feature. The first-level noise filtering component removes these high-frequency noises by smoothing or median filtering. After the first-level processing, the feature data enters the intermediate-level noise filtering component. These components may be designed for more complex noise patterns, such as periodic noise, baseline drift, etc. For example, due to the slow change of ambient temperature, the baseline of the sensor may gradually drift, resulting in a long-term, slow change trend in the fused topological feature. The intermediate noise filtering component can identify and correct this baseline drift through trend analysis or differential processing. In the last level of noise filtering components, the server further processes the low-frequency noise or outliers that may remain. These outliers may be caused by sensor aging, calibration errors, or other unknown factors. The last level component can use more complex statistical models or machine learning algorithms to identify and remove these outliers to ensure the accuracy and reliability of the fused topology features. After being processed by multiple levels of noise filtering components, the server finally outputs the optimized fused topology features. These features are cleaner and more accurate, and can better reflect the essential characteristics of the abnormal conditions of the target locomotive components. Assume that the fused topology features obtained by the server contain high-frequency noise caused by electromagnetic interference, baseline drift caused by ambient temperature changes, and outliers caused by sensor aging. By passing through multiple cascaded noise filtering components in sequence, the server can gradually remove these noises and outliers, and finally obtain a set of optimized fused topology features, which will be used for subsequent abnormal condition inference and analysis.

[0125] In the embodiment of the present invention, before performing noise filtering on the fused topological features based on the noise filtering unit to obtain the optimized features of the abnormal conditions of the target locomotive components, the following implementation methods may also be provided.

[0126] Acquire a raw noise filtering unit including a plurality of raw noise filtering components, a target locomotive component abnormality sample, and a known locomotive component abnormality feature of a known locomotive component abnormality;

[0127] Extracting target description content from the target locomotive component abnormality sample to obtain sample target description content, and extracting sample hypergraphs corresponding to each sample target description content from the hypergraph;

[0128] Extracting topological features from each of the sample hypergraphs to obtain sample topological features of each of the sample hypergraphs, and fusing features of each of the sample topological features to obtain fused sample features of the target locomotive component abnormality sample;

[0129] Based on each of the original noise filtering components of the original noise filtering unit, respectively, noise filtering is performed on the fused sample features to obtain optimized sample features corresponding to each of the original noise filtering components;

[0130] The cost parameter of the original noise filtering unit is determined by combining each of the optimized sample features and the known locomotive component abnormality features, and based on the cost parameter, a training process is performed on the original noise filtering unit to obtain the noise filtering unit.

[0131] In an embodiment of the present invention, exemplarily, the server obtains multiple original noise filtering components from the database, and these components will form an original noise filtering unit. At the same time, the server collects samples of abnormal conditions of target locomotive components and feature data of known abnormal conditions of locomotive components. The server downloads five different original noise filtering components from the storage system, which are designed for different types of noise. The server also selects 100 samples of abnormal conditions of target locomotive components and known abnormal condition features corresponding to these samples from the historical data of locomotive operation. The server extracts the target description content of each target locomotive component abnormal condition sample to obtain the sample target description content. Subsequently, the server extracts the sample hypergraph corresponding to each sample target description content from the hypergraph. For a sample of abnormal condition of a brake system, the server extracts its target description content, such as "prolonged brake response time". Then, the server finds the sample hypergraph corresponding to the description content from the hypergraph, and this hypergraph shows in detail various associations and features of the brake system. The server extracts topological features of each sample hypergraph to obtain sample topological features of each sample hypergraph. These topological features are then fused to form fused sample features of the target locomotive component abnormal condition sample. For a sample hypergraph of an engine with excessively high temperature, the server extracts multiple topological features including the reading of the temperature sensor and the status of the cooling system. These features are fused into a comprehensive fused sample feature for subsequent analysis and filtering. The server uses each original noise filter component of the original noise filter unit in turn to filter the fused sample feature. After filtering, the server obtains the optimized sample feature corresponding to each original noise filter component. The first noise filter component is used to remove the high-frequency noise in the fused sample feature. The second component is used to deal with the baseline drift problem caused by sensor aging. After filtering by all components, the server obtains a set of optimized sample features. The server combines each optimized sample feature with the known locomotive component abnormality features to determine the cost parameters of the original noise filter unit. Based on these cost parameters, the server performs a training process on the original noise filter unit and finally obtains the optimized noise filter unit. The server calculates the cost parameters, such as the mean square error, by comparing the differences between the optimized sample features and the known abnormality features. Based on these cost parameters, the server adjusts the parameters and configurations in the original noise filter unit to make it more suitable for the noise filtering needs of the target locomotive component abnormality. After multiple iterations of training, the server obtained an efficient noise filtering unit, which provided an accurate data basis for subsequent fault diagnosis.

[0132] In an embodiment of the present invention, the original noise filtering unit includes multiple original noise filtering components, the characteristic dimension of the known locomotive component abnormality feature is consistent with the characteristic dimension of each of the optimized sample features, and the cost parameter of the original noise filtering unit is determined by combining each of the optimized sample features and the known locomotive component abnormality feature, which can be implemented through the following example.

[0133] Determining a first cost parameter of the original noise filtering unit based on each of the optimized sample features;

[0134] Calculate the feature difference between the optimized sample feature corresponding to the last original noise filtering component and the known locomotive component abnormality feature to obtain a reference feature, and determine the feature length of the reference feature as the second cost parameter of the original noise filtering unit;

[0135] The first cost parameter and the second cost parameter are accumulated to obtain the cost parameter.

[0136] In the embodiment of the present invention, exemplarily, the server has obtained the original noise filtering unit (including multiple original noise filtering components), the fusion sample features of the target locomotive component abnormality sample, and the optimized sample features after being processed by the original noise filtering unit through the previous steps. At the same time, the server also has known locomotive component abnormality features consistent with the dimensions of these optimized sample features. The server first determines the first cost parameter of the original noise filtering unit based on each optimized sample feature. This parameter is usually used to measure the optimization effect of the noise filtering unit on the sample feature. The server can use indicators such as mean square error and signal-to-noise ratio to quantify the difference between each optimized sample feature and the original sample feature, thereby obtaining the first cost parameter. The server calculates the mean square error between 100 optimized sample features and the original sample features, and uses the average value of these errors as the first cost parameter. This parameter reflects the overall optimization performance of the noise filtering unit. The server then calculates the feature difference between the optimized sample features corresponding to the last original noise filtering component and the known locomotive component abnormality features. This usually involves calculating the distance or similarity between two feature vectors. Through this difference calculation, the server obtains a benchmark feature, which reflects the gap between the optimized sample feature and the ideal situation (i.e., the known locomotive component abnormal situation feature) after the complete noise filtering process. The server determines the feature length of the benchmark feature (such as the complement of the Euclidean distance or cosine similarity) as the second cost parameter of the original noise filtering unit. The server uses the Euclidean distance to calculate the difference between the optimized sample feature corresponding to the last original noise filtering component and the known locomotive component abnormal situation feature. This difference value is used as the feature length of the benchmark feature and is determined as the second cost parameter. This parameter reflects the performance of the noise filtering unit in approaching the ideal situation. Finally, the server performs cumulative processing on the first cost parameter and the second cost parameter to obtain a total cost parameter. This total cost parameter comprehensively considers the performance of the noise filtering unit in terms of overall optimization performance and approaching the ideal situation. The server adds the first cost parameter (the average value of the mean square error) and the second cost parameter (the feature length of the benchmark feature) calculated previously to obtain a total cost parameter. This parameter is used as a basis for subsequent optimization or evaluation of the performance of the noise filtering unit.

[0137] In the embodiment of the present invention, the determining of the first cost parameter of the original noise filtering unit based on each of the optimized sample features may be implemented through the following examples.

[0138] Obtaining feature difference values ​​of optimized sample features corresponding to the first original noise filter component and optimized sample features corresponding to the second original noise filter component to obtain a feature difference vector corresponding to the second original noise filter component, and determining the feature length of the feature difference vector as a benchmark cost parameter corresponding to the second original noise filter component; the second original noise filter component is a subsequent original noise filter component of the first original noise filter component;

[0139] Each of the reference cost parameters is combined and calculated to obtain the first cost parameter.

[0140] In an embodiment of the present invention, illustratively, the server first selects the first original noise filter component and the second original noise filter component in the original noise filter unit. Here, the second original noise filter component is a subsequent component of the first original noise filter component. The server obtains feature difference values ​​of the optimized sample features corresponding to the first original noise filter component and the optimized sample features corresponding to the second original noise filter component. This usually involves element-by-element difference calculation of two feature vectors.

[0141] Through this calculation, the server obtains a feature difference vector, which reflects the difference in optimized sample features between two adjacent noise filter components. Assume that the server has two adjacent original noise filter components A and B, where A is the first original noise filter component and B is the second original noise filter component. The server calculates the difference between the optimized sample features after processing A and B, and obtains a feature difference vector. This vector shows which features have changed significantly from A to B. The server then calculates the feature length of the feature difference vector. This can be achieved by calculating the Euclidean distance, L2 norm, or other suitable metrics of the vector. This feature length is determined as the benchmark cost parameter corresponding to the second original noise filter component. It quantifies the degree of change in the optimized sample features from the first original noise filter component to the second original noise filter component. The server calculates the L2 norm of the feature difference vector and obtains a specific value. This value reflects the overall change in sample features during the optimization process from noise filter component A to B. This L2 norm value is determined as the benchmark cost parameter of noise filter component B. The server repeats the above process for each pair of adjacent noise filter components in the original noise filter unit, thereby obtaining a series of benchmark cost parameters. Finally, the server performs a combined calculation on these benchmark cost parameters, such as summing, weighted averaging, etc., to obtain a first cost parameter. This parameter comprehensively reflects the overall performance of the entire noise filtering unit in the process of optimizing sample features. The server calculates the benchmark cost parameters between all adjacent noise filtering components in the original noise filtering unit. Then, the server sums these benchmark cost parameters to obtain a total first cost parameter. This parameter is used to evaluate the effect of the entire noise filtering unit in optimizing the sample features of abnormal conditions of locomotive components.

[0142] Please refer to Figure 2 , Figure 2 This is a schematic diagram of the structure of a handheld voiceprint collection terminal 30 provided in an embodiment of the present invention.

[0143] In the embodiment of the present invention, the handheld voiceprint collection terminal 30 includes a microphone array 301 and a sensor 302;

[0144] The microphone array 301 and the sensor 302 are used to collect the current device voiceprint and the environmental parameters of the above solution respectively. In the embodiment of the present invention, the sensor 302 can be a camera. In addition, a temperature sensor, a humidity sensor, etc. can also be configured to collect environmental parameters, which is not limited here.

[0145] Please refer to Figure 3 , Figure 3 A schematic diagram of the exploded structure of the handheld voiceprint collection terminal 30 provided in an embodiment of the present invention.

[0146] In the embodiment of the present invention, the handheld voiceprint collection terminal 30 includes not only a microphone array 301 and a sensor 302, but also a communication module 303 for communicating and interacting with a server; a display screen 304 for displaying the device status, the content of the collected data, and the current locomotive equipment maintenance plan sent back by the server; a battery 305 for independently providing power to the handheld voiceprint collection terminal 30; and a grip 306 for providing a handheld position. In addition, the handheld voiceprint collection terminal 30 may also include a fill light, a switch, a heat sink, an earphone jack, and other structures, which will not be described in detail here.

[0147] Please refer to Figure 4 , Figure 4 A locomotive equipment maintenance system 110 based on a handheld voiceprint collection terminal provided in an embodiment of the present invention includes:

[0148] The acquisition module 1101 is used to acquire the current equipment voiceprint and environmental parameters collected by the handheld voiceprint collection terminal at a preset locomotive equipment maintenance position;

[0149] Processing module 1102 is used to input the current device voiceprint into the voiceprint abnormal type recognition model to obtain the current abnormal voiceprint recognition result corresponding to the current device voiceprint; generate the current locomotive device abnormality description according to the current abnormal voiceprint recognition result and the environmental parameters; and obtain the inferred locomotive component abnormality according to the current locomotive device abnormality description;

[0150] The maintenance module 1103 is used to obtain a candidate locomotive equipment maintenance plan corresponding to the inferred locomotive component abnormality from a preset maintenance database as a current locomotive equipment maintenance plan.

[0151] It should be noted that the implementation principle of the locomotive equipment maintenance system 110 based on the handheld voiceprint collection terminal can refer to the implementation principle of the locomotive equipment maintenance method based on the handheld voiceprint collection terminal, which will not be repeated here. It should be understood that the division of the various modules of the above device is only a division of logical functions. In actual implementation, they can be fully or partially integrated into a physical entity, or physically separated. And these modules can all be implemented in the form of software called by processing elements; they can also be all implemented in the form of hardware; some modules can also be implemented in the form of software called by processing elements, and some modules can be implemented in the form of hardware. For example, the locomotive equipment maintenance system 110 based on the handheld voiceprint collection terminal can be a separately established processing element, or it can be integrated in a chip of the above device. In addition, it can also be stored in the memory of the above device in the form of program code, and a processing element of the above device can call and execute the functions of the locomotive equipment maintenance system 110 based on the handheld voiceprint collection terminal. The implementation of other modules is similar. In addition, all or part of these modules can be integrated together, or they can be implemented independently. The processing element described here can be an integrated circuit with signal processing capabilities. In the implementation process, each step of the above method or each module above can be completed by an integrated logic circuit of hardware in a processor element or an instruction in software form.

[0152] For example, the above modules may be one or more integrated circuits configured to implement the above methods, such as one or more application specific integrated circuits (ASIC), or one or more microprocessors (digital signal processors, DSP), or one or more field programmable gate arrays (FPGA), etc. For another example, when a module above is implemented in the form of a processing element scheduling program code, the processing element may be a general-purpose processor, such as a central processing unit (CPU) or other processor that can call program code. For another example, these modules may be integrated together and implemented in the form of a system-on-a-chip (SOC).

[0153] 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 aforementioned locomotive equipment maintenance system 110 based on a handheld voiceprint collection terminal. Figure 5 As shown, Figure 5The computer device 100 provided in the embodiment of the present invention is a structural block diagram. The computer device 100 includes a locomotive equipment maintenance system 110 based on a handheld voiceprint collection terminal, a memory 111, a processor 112 and a communication unit 113.

[0154] In order to realize data transmission or interaction, the memory 111, the processor 112 and the communication unit 113 are electrically connected to each other directly or indirectly. For example, the electrical connection between these elements can be realized through one or more communication buses or signal lines. The locomotive equipment maintenance system 110 based on the handheld voiceprint collection terminal includes at least one software function module that can be stored in the memory 111 in the form of software or firmware or solidified in the operating system (OS) of the computer device 100. The processor 112 is used to execute the locomotive equipment maintenance system 110 based on the handheld voiceprint collection terminal stored in the memory 111, such as the software function modules and computer programs included in the locomotive equipment maintenance system 110 based on the handheld voiceprint collection terminal.

[0155] 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 locomotive equipment maintenance method based on a handheld voiceprint collection terminal, characterized in that: include: Obtain the current equipment voiceprint and environmental parameters collected by the handheld voiceprint collection terminal at the preset locomotive equipment maintenance location; Inputting the current device voiceprint into a voiceprint abnormal type recognition model to obtain a current abnormal voiceprint recognition result corresponding to the current device voiceprint; Generate a current locomotive equipment abnormality description according to the current abnormal voiceprint recognition result and the environmental parameters; Inferring the abnormality of locomotive components according to the current locomotive equipment abnormality description; A candidate locomotive equipment maintenance plan corresponding to the inferred locomotive component abnormality is obtained from a preset maintenance database as a current locomotive equipment maintenance plan.

2. The method according to claim 1, characterized in that The voiceprint abnormality type recognition model is obtained by the following methods, including: Segmenting the first device voiceprint instance, and performing a feature extraction operation on each segmented voiceprint segment to obtain an original voiceprint feature data set, wherein the original voiceprint feature data set contains an original voiceprint feature corresponding to each voiceprint segment, and the first device voiceprint instance is a device voiceprint instance of a pre-configured category but not configured with an abnormal type; Reorganizing and arranging the original voiceprint features in the original voiceprint feature data set to obtain a first voiceprint feature data set and a second voiceprint feature data set, wherein the first voiceprint feature in the first voiceprint feature data set and the second voiceprint feature in the second voiceprint feature data set correspond to different reorganization and arrangement strategies; Pre-training a voiceprint abnormality type recognition model according to the first voiceprint feature data set and the second voiceprint feature data set, wherein the voiceprint abnormality type recognition model is used to recognize abnormal situations in voiceprints; The pre-trained voiceprint abnormality type recognition model is tuned according to the second device voiceprint instance, where the second device voiceprint instance is a device voiceprint instance of a pre-configured category.

3. The method according to claim 2, characterized in that The reorganizing and arranging the original voiceprint features in the original voiceprint feature data set to obtain a first voiceprint feature data set and a second voiceprint feature data set includes: Adjusting the feature arrangement order of the original voiceprint features in the original voiceprint feature data set to obtain a first original voiceprint feature data set and a second original voiceprint feature data set, wherein the arrangement order of the original voiceprint features in the first original voiceprint feature data set and the second original voiceprint feature data set are different; Determining a vector dimension according to a segmentation strategy of the voiceprint instance of the first device; According to the vector dimension, the original voiceprint features in the first original voiceprint feature data set are sequenced to obtain the first voiceprint feature vector; According to the vector dimension, reorganizing the original voiceprint features in the second original voiceprint feature data set to obtain the second voiceprint feature vector; Selecting a plurality of continuous original voiceprint features in the first voiceprint feature vector by a continuous feature selector; Performing feature concatenation on the multiple original voiceprint features to obtain a first integrated voiceprint feature, or performing feature integration on the multiple original voiceprint features to obtain the first integrated voiceprint feature; Performing feature concatenation on the multiple original voiceprint features to obtain a second integrated voiceprint feature, or performing feature integration on the multiple original voiceprint features to obtain the second integrated voiceprint feature; Performing feature conversion on multiple groups of the first integrated voiceprint features to obtain the first voiceprint feature data set, wherein the multiple groups of the first integrated voiceprint features are obtained by shifting the continuous feature selector; Selecting a plurality of continuous original voiceprint features in the second voiceprint feature vector by a continuous feature selector; Performing feature integration on the multiple original voiceprint features to obtain a second integrated voiceprint feature; Feature conversion is performed on multiple groups of the second integrated voiceprint features to obtain the second voiceprint feature data set, and the multiple groups of the second integrated voiceprint features are obtained by shifting the continuous feature selector.

4. The method according to claim 2, characterized in that: After the original voiceprint features in the original voiceprint feature data set are reorganized and arranged to obtain the first voiceprint feature data set and the second voiceprint feature data set, the method includes: According to the first voiceprint feature data set, cyclically performing reorganization and arrangement at least once to obtain a third voiceprint feature data set; According to the second voiceprint feature data set, cyclically performing reorganization and arrangement at least once to obtain a fourth voiceprint feature data set; A voiceprint abnormality type recognition model is pre-trained according to the third voiceprint feature data set and the fourth voiceprint feature data set.

5. The method according to claim 2, characterized in that: The pre-training of the voiceprint abnormality type recognition model according to the first voiceprint feature data set and the second voiceprint feature data set includes: Loading the first voiceprint feature data set into the dynamic network branch of the voiceprint abnormal type recognition model to obtain a first abnormal voiceprint recognition result; Loading the second voiceprint feature data set into the static network branch of the voiceprint abnormal type recognition model to obtain a second abnormal voiceprint recognition result; Determine a matching degree error between the first abnormal voiceprint recognition result and the second abnormal voiceprint recognition result; According to the matching degree error, adjusting the network parameters of the dynamic network branch by error feedback; According to the network parameters of the dynamic network branch that have completed training, the network parameters of the static network branch are adjusted through model integration.

6. The method according to claim 5, characterized in that The step of optimizing the pre-trained voiceprint abnormality type recognition model according to the second device voiceprint instance includes: Loading the second device voiceprint instance to the static network branch of the voiceprint abnormal type recognition model to obtain a sample abnormal voiceprint recognition result; According to the sample abnormal voiceprint recognition result and the sample abnormal type configuration corresponding to the second device voiceprint instance, the network parameters of the static network branch are adjusted through error feedback.

7. The method according to claim 1, characterized in that The inferring the abnormality of the locomotive component according to the current locomotive equipment abnormality description includes: Acquire the abnormal condition of the locomotive component to be analyzed, and identify the abnormal description of the current locomotive equipment for the abnormal condition of the locomotive component to be analyzed, so as to obtain the abnormal description information of the current locomotive equipment for the abnormal condition of the locomotive component to be analyzed; When the current locomotive equipment abnormality description information indicates that the locomotive component abnormality to be analyzed includes the current locomotive equipment abnormality description, determining the locomotive component abnormality to be analyzed as the target locomotive component abnormality; Performing semantic analysis on the abnormal condition of the target locomotive component to obtain multiple key descriptions of the abnormal condition of the target locomotive component; For each of the key descriptions, when the key description includes the current locomotive equipment abnormality description, determining the key description as the target description content; Obtain a hypergraph for representing the association relationship between multiple reference description contents, and compare the target description content with each of the reference description contents in the hypergraph respectively to obtain a comparison result, wherein the comparison result is used to represent whether the hypergraph contains reference description contents that are consistent with the target description contents; When the comparison result indicates that the hypergraph contains reference description content that is consistent with the target description content, the reference description content that is consistent with the target description content is determined as the target reference description content corresponding to the target description content; When the comparison result indicates that the hypergraph does not contain reference description content that is consistent with the target description content, obtaining a matching degree between the target description content and each of the reference description contents; Determine the reference description content whose matching degree exceeds the matching degree threshold as the target reference description content corresponding to the target description content; Determine the entity corresponding to the target reference description content in the hypergraph as the target entity corresponding to the target description content, and determine the entity in the hypergraph that has a connection relationship with the target entity as the reference entity corresponding to the target description content; Determine the area in the hypergraph formed by the target entity and the reference entity as a local hypergraph corresponding to the target description content; There is an association relationship between the entity in the hypergraph and the reference description content, the local hypergraph includes a target entity, and the matching degree between the reference description content corresponding to the target entity and the corresponding target description content exceeds a matching degree threshold; Extracting topological features from each of the local hypergraphs to obtain topological features of each of the local hypergraphs, and fusing features of each of the topological features to obtain fused topological features of the abnormal conditions of the target locomotive components; Acquire a raw noise filtering unit including a plurality of raw noise filtering components, a target locomotive component abnormality sample, and a known locomotive component abnormality feature of a known locomotive component abnormality; Extracting target description content from the target locomotive component abnormality sample to obtain sample target description content, and extracting sample hypergraphs corresponding to each sample target description content from the hypergraph; Extracting topological features from each of the sample hypergraphs to obtain sample topological features of each of the sample hypergraphs, and fusing features of each of the sample topological features to obtain fused sample features of the target locomotive component abnormality sample; Based on each of the original noise filtering components of the original noise filtering unit, the fused sample features are respectively subjected to noise filtering to obtain optimized sample features corresponding to each of the original noise filtering components; the original noise filtering unit includes a plurality of original noise filtering components, and the feature dimension of the known locomotive component abnormality feature is consistent with that of each of the optimized sample features; Obtaining feature difference values ​​of optimized sample features corresponding to the first original noise filter component and optimized sample features corresponding to the second original noise filter component to obtain a feature difference vector corresponding to the second original noise filter component, and determining the feature length of the feature difference vector as a benchmark cost parameter corresponding to the second original noise filter component; the second original noise filter component is a subsequent original noise filter component of the first original noise filter component; Performing a combined calculation on each of the reference cost parameters to obtain a first cost parameter; Calculate the feature difference between the optimized sample feature corresponding to the last original noise filtering component and the known locomotive component abnormality feature to obtain a reference feature, and determine the feature length of the reference feature as the second cost parameter of the original noise filtering unit; The first cost parameter and the second cost parameter are accumulated to obtain the cost parameter, and based on the cost parameter, a training process is performed on the original noise filtering unit to obtain a noise filtering unit; the noise filtering unit includes a plurality of noise filtering components; Performing noise filtering on the fused topological features through a plurality of cascaded noise filtering components in sequence to obtain optimized features of the abnormal conditions of the target locomotive components; Based on the abnormal situation inference unit, according to the optimization feature, the abnormal situation of the target locomotive component is inferred to obtain the inferred locomotive component abnormal situation corresponding to the target locomotive component abnormal situation, and the current locomotive equipment abnormality description of the inferred locomotive component abnormal situation is consistent with the current locomotive equipment abnormality description of the target locomotive component abnormal situation.

8. The method according to claim 7, characterized in that After identifying the current locomotive equipment abnormality description of the locomotive component abnormality to be analyzed and obtaining the current locomotive equipment abnormality description information of the locomotive component abnormality to be analyzed, the method further includes: When the current locomotive equipment abnormality description information indicates that the abnormality of the locomotive component to be analyzed does not include the current locomotive equipment abnormality description, respectively extracting the reference hypergraphs corresponding to each of the key descriptions from the hypergraph, performing topological feature extraction on each of the reference hypergraphs to obtain reference topological features of each of the reference hypergraphs, and performing feature fusion on each of the reference topological features to obtain reference fused topological features of the abnormality of the locomotive component to be analyzed; Based on the reference fusion topological features, abnormality inference is performed on the abnormality of the locomotive component to be analyzed to obtain a target inferred locomotive component abnormality corresponding to the abnormality of the locomotive component to be analyzed; The target inferred locomotive component abnormality includes the current locomotive equipment abnormality description, and the target inferred locomotive component abnormality includes the locomotive component abnormality to be analyzed.

9. A handheld voiceprint collection terminal, characterized in that: Includes microphone array and sensor; The microphone array and the sensor are used to collect the current device voiceprint and the environmental parameters in claim 1 respectively.

10. A locomotive equipment maintenance system based on a handheld voiceprint collection terminal, characterized in that: include: An acquisition module, used to acquire the current equipment voiceprint and environmental parameters collected by the handheld voiceprint collection terminal at a preset locomotive equipment maintenance location; A processing module, configured to input the current device voiceprint into a voiceprint abnormality type recognition model to obtain a current abnormal voiceprint recognition result corresponding to the current device voiceprint; and generate a current locomotive device abnormality description according to the current abnormal voiceprint recognition result and the environmental parameters; Inferring the abnormality of locomotive components according to the current locomotive equipment abnormality description; The maintenance module is used to obtain a candidate locomotive equipment maintenance plan corresponding to the inferred locomotive component abnormality from a preset maintenance database as a current locomotive equipment maintenance plan.

Images