1580 results about "Spectrogram" patented technology

The invention relates to the field of enterprise risk early warning analysis, in particular to an intelligent risk early warning method and system based on multi-dimensional data analysis. The method comprises the following steps: acquiring a multi-dimensional enterprise data stream, performing heterogeneous index information analysis and logic hierarchy reconstruction, and constructing an enterprise running state sensing map; performing multi-index local fluctuation amplitude calculation on the enterprise operation state sensing map, and performing dynamic disturbance feature mining to construct a risk disturbance spectrogram; performing deep semantic analysis and node state abrupt change feature analysis on the risk disturbance spectrogram to generate a dynamic transition type of an abrupt change node; and performing full-period time sequence tracing on the risk disturbance spectrogram, performing potential risk node prediction based on the dynamic transition type, and identifying other potential risk propagation ports. Through accurate and efficient enterprise risk perception, the risk intervention decision is made in advance, and the anti-risk capability and the operation stability of the enterprise are improved.

A method for early diagnosis of Parkinson's disease based on multimodal deep learning is provided. Audio-visual data of a to-be-diagnosed subject while performing a speech task is acquired. The audio-visual data are preprocessed to extract a plurality of audio segments and a plurality of video segments. A face image sequence is extracted from each of the plurality of video segments. A Mel-spectrogram of each of the plurality of audio segments is calculated. The face image sequence and the Mel-spectrogram are input into a multimodal deep learning model to output a classification result for Parkinson's disease early diagnosis of the to-be-diagnosed subject. A system for early diagnosis of Parkinson's disease based on multimodal deep learning is also provided.

The invention relates to the technical field of audio noise reduction, and discloses a Bluetooth headset environment noise adaptive adjustment method and system, and the method comprises the steps: collecting multi-band sound data of headset environment noise through a preset acoustic sensor array, generating a noise spectrogram, extracting a noise masking threshold through time-frequency joint analysis, and collecting a headset audio signal; next, noise suppression is carried out on the audio signal by using a multi-order digital filter bank to obtain a primary noise reduction signal, the primary noise reduction signal is optimized into an optimized noise reduction signal based on a signal-to-noise ratio, and the frequency response matching degree of the optimized noise reduction signal and the original audio signal is calculated; then, performing phase correction and amplitude-frequency equalization processing according to the frequency response matching degree to obtain an audio equalization parameter; and finally, monitoring residual noise energy based on the parameter, generating a noise regulation and control instruction, driving adaptive sampling, and generating an adaptive noise reduction scheme. According to the invention, the audio noise reduction effect of the Bluetooth earphone can be improved.

A system configured to improve audio processing by performing dereverberation and noise reduction during a communication session. In some examples, the system may include a deep neural network (DNN) configured to perform speech enhancement, which is located after an Acoustic Echo Cancellation (AEC) component. For example, the DNN may process isolated audio data output by the AEC component to jointly mitigate additive noise and reverberation. In other examples, the system may include a DNN configured to perform acoustic interference cancellation, which may jointly mitigate additive noise, reverberation, and residual echo, removing the need to perform residual echo suppression processing. The DNN is configured to process complex-valued spectrograms corresponding to the isolated audio data and / or estimated echo data generated by the AEC component.

The invention discloses a precise interference avoidance method and device in a radio system, and relates to the field of signal processing, and the method comprises the steps: constructing a sensing matrix through sensing node data, and capturing a transient interference signal through aperiodic scanning; performing tensor decomposition on the signal data to extract time domain, frequency domain, space domain and modulation domain features, constructing a dual-mode spectrum analysis model, reconstructing an instantaneous spectrogram by using compressed sensing, and predicting an interference mode through LSTM; after the instantaneous spectrogram and the predicted interference graph are fused, threat assessment is carried out through a multi-stage interference classification model; according to the interference category and the threat level, beam forming is optimized, adaptive null is generated, a power density optimization model is constructed, and the transmitting power is dynamically adjusted; an anti-interference frequency hopping sequence is generated based on a chaotic mapping algorithm, and spectrum camouflage and tracking interference resistance are realized. The method has the advantages that accurate identification and dynamic avoidance of interference are realized through multi-dimensional perception, intelligent prediction and adaptive beam forming, and the interference avoidance capability of a wireless system is improved.

The invention discloses a water supply network sound leakage monitoring method based on multi-modal data fusion and multi-task learning, and the method comprises the steps: inputting an original sound signal, and carrying out the labeling of a label; converting the original sound signal to obtain a corresponding logarithmic Mel spectrogram; performing feature extraction on the original sound signal to construct a time-frequency feature set; forming a data set by the original sound signal, the logarithmic Mel spectrogram, the label and the time-frequency feature set; constructing a multistage task prediction network based on a residual convolutional neural network framework; performing multi-task training on the multi-level task prediction network by using the data set to obtain a water supply network sound leakage monitoring model; and inputting the original sound signals of the collection points into the water supply network sound leakage monitoring model to obtain a prediction result. The invention further provides a water supply network sound leakage monitoring device. According to the method provided by the invention, the leakage detection aspect is promoted to more comprehensive sound leakage monitoring application through a data-driven acoustic detection technology.

The invention relates to the technical field of carrier communication, in particular to a multimode power line carrier communication method and a carrier communication module, and the method comprises the steps: determining a frequency coincidence degree through analyzing the coincidence degree of a frequency corresponding to a positive number amplitude value on a spectrogram of each modal component in a carrier signal and a signal frequency range of the carrier signal, and determining a difference index by analyzing the difference between each characteristic modal component and all modal components of the noise reference signal in frequency, analyzing the centralized distribution condition of the carrier signal in energy, and determining a step length adjustment value in combination with the signal interference degree. And carrying out noise reduction processing on the carrier signal in the power line carrier communication process at the current moment. The invention aims to eliminate noise interference in carrier signals and improve the reliability of power line carrier communication.

The invention discloses a method for identifying a multi-model fusion signal modulation mode based on a time-frequency diagram, and belongs to the field of signal processing. Preprocessing the original signal to obtain a signal sample; constructing a data set by using the normalized signal samples; constructing a fusion model; training a modulation identification module in the fusion model; performing modulation mode identification on an input unknown signal by using the fusion model to obtain a preliminary identification result; and performing comprehensive judgment on the preliminary recognition result of the fusion model by using a comprehensive judgment device to obtain a final recognition result. According to the invention, by combining a plurality of time-frequency analysis methods, the time-frequency domain feature information of the signal is fully extracted; the reliability and robustness of a final decision are improved by adopting a multi-model fusion framework, and the modulation recognition performance of the system under the condition of a low signal-to-noise ratio is improved.

An electronic device is provided. The electronic device includes a memory, and at least one processor electrically connected to the memory, wherein the at least one processor is configured to obtain a video including an image and an audio, obtain information on at least one object included in the image from the image, obtain a visual feature of the at least one object, based on the image and the information on the at least one object, obtain a spectrogram of the audio, obtain an audio feature of the at least one object from the spectrogram of the audio, combine the visual feature and the audio feature, obtain, based on the combined visual feature and audio feature, information on a position of the at least one object the information indicating the position of the at least one object in the image, obtain an audio part corresponding to the at least one object in the audio, based on the combined visual feature and audio feature, and store, in the memory, the information on the position of the at least one object and the audio part corresponding to the at least one object.

The invention discloses a water turbine fault classification diagnosis method based on multi-modal fusion and meta-learning. The method comprises the following steps: step 1, respectively extracting time domain features and Mel-language spectrogram features of monitoring noise signals of a water turbine through a feature extraction module; step 2, performing cross-modal attention mechanism fusion on the time domain features and the voiceprint features through a multi-modal fusion module, and adjusting a fusion weight based on a dynamic weight distribution mechanism; and step 3, performing small sample training optimization on the fused features through a meta-learning module, and improving the classification capability of the model for new fault types in combination with a Triplet Loss-KNN algorithm and twin network pre-training. The fault diagnosis method based on the time domain-voiceprint fusion network and meta learning has the advantages of being rapid in diagnosis, accurate in classification, high in generalization ability and the like, and the fault diagnosis precision of the water turbine based on noise signals can be effectively improved.

The invention provides a multi-language cross-culture communication assisting method and system based on a large model. The method comprises the following steps: receiving a source language audio stream during a call, calling a multi-language sound frequency harmonic modulation feature library to extract fundamental frequency harmonic intensity distribution and tone turning features, and generating a cultural acoustic fingerprint vector; based on the vector, controlling a microphone array phase difference, directionally enhancing a fundamental frequency harmonic component of a speaker and suppressing noise, and outputting a high signal-to-noise ratio spectrogram; analyzing the pronunciation rhythm and tone turning characteristics of the spectrogram, capturing the pitch jump and duration of the syllable boundary, and generating an acoustic culture label; associating the spectrogram with a target semantic library, matching harmonic distribution and a cultural context rule based on a large model, and outputting a cultural interpretation prompt containing an ambiguity resolution suggestion; and generating a calibration result according to the acoustic tag and the semantic prompt, and overlapping the dynamic floating subtitles to the face area of the speaker in the video conference picture. According to the invention, cultural tone ambiguity in multi-language communication is eliminated.

Systems, methods, and computer-readable media are disclosed for systems and methods multimodal indexing of video using machine learning. An example method may include deceiving, by a video encoder of an audio-video transformer neural network comprising one or more computer processors coupled to memory, a first frame and a second frame associated with a first segment of a video. The example method may also include receiving, by an audio encoder of the audio-video transformer neural network, an audio spectrogram comprising first audio data associated with the first segment of the video. generating, by the video encoder, a first video embedding. The example method may also include generating, by the audio encoder, a first audio embedding. The example method may also include determining a fusion of the first video embedding and the first audio embedding using a multimodal bottleneck token. The example method may also include determining an output including the first video embedding and the first audio embedding. The example method may also include determining a classification of the first portion of the video based on the output.

The invention discloses a multi-mode counterfeit video detection method based on a multi-head addition cross attention mechanism, and belongs to the technical field of video counterfeit detection. The method comprises the following steps: preprocessing a video stream, decomposing a single-frame positioning face, and extracting an audio to generate a Mel spectrogram slice; the 3D convolutional network extracts video spatio-temporal features and motion differences, and the filter bank extracts audio features in combination with the residual network; audio features are mapped to a video alignment space through asymmetric projection, the video features are subjected to bidirectional interaction with an audio input multi-head addition cross attention module after being subjected to time sequence coding, and audio dominant and video dominant features are generated and are cascaded and fused with original features; and constructing cross-modal similarity loss constraint feature distribution, and fusing feature dynamic weighting and time sequence compression to output four classification probabilities of audio-visual double true, audio-visual double pseudo, video pseudo-audio true and video pseudo-audio pseudo. The multi-mode counterfeiting recognition precision is improved, and texture abnormity and audio and video mismatch features are captured.

The invention relates to the technical field of speech synthesis, can be applied to business scenes of medical health, financial science and technology and the like, and discloses a hierarchical emotional speech generation method, which comprises the following steps: acquiring an input text and an emotional speech sample, extracting a text embedding feature from the input text, extracting a Mel spectrum feature from the emotional speech sample, and obtaining a text embedding feature; the method comprises the following steps: extracting phoneme-level, word-level and statement-level sentiment distribution characteristics through a hierarchical sentiment distribution extraction module, carrying out time dimension alignment, generating a multi-level sentiment guidance matrix, and inputting the multi-level sentiment guidance matrix and text embedding characteristics into a sentiment synthesis module to generate a target Mel spectrum; and finally, converting the target Mel spectrum into target voice through a vocoder. The emotion control is expanded from the statement level to the fine-grained level of phonemes, words and the like, and the multi-level emotion guidance matrix is combined for generation, so that the generated speech is finer and more abundant in emotion expression and conforms to the context, and the naturalness of speech generation and the accuracy of emotion transmission are improved.

The present invention relates to a computer implemented method for monitoring the blade (210) of a wind turbine (200), the method comprising fixing a sound acquisition device (110) to an outer surface of a tower (220) of the wind turbine (200), wherein the sound acquisition device (110) comprises a microphone, a sound encoder and a communication unit, acquiring a sound sample of sound emitted by the blade (210) during operation of the wind turbine (200), encoding the sound sample into a raw audio file, sending the raw audio file to a remote computing device (120), preprocessing the raw audio file, generating a spectrogram of the pre-processed audio file, detecting an anomaly by applying an image analysis machine learning model to the spectrogram, classifying the anomaly to a damage type by applying a damage type feature analysis machine learning model to frequency domain features of the audio file at the time and frequency of the anomaly, determining a damage location by applying a damage location feature analysis machine learning model to frequency domain features of the audio file at the time and frequency of the, and tracking of a damage severity by extracting audio features of the audio file at the time and frequency of the anomaly indicative of the damage severity depending on the damage type.

The invention discloses a pipeline defect identification method and device, a storage medium and equipment, and belongs to the technical field of pipeline defect detection. The method comprises the following steps: S1, acquiring a time sequence sound signal acquired by an acoustic sensor arranged on the inner wall and / or the outer wall of a pipeline; s2, performing spectrogram extraction on the time sequence sound signal to obtain a Mel spectrogram and a LOFAR spectrogram; and S3, performing feature extraction on the Mel spectrogram, the LOFAR spectrogram and the time sequence sound signal through a deep learning model of multi-modal signal fusion to obtain a feature vector containing spectrogram features and time sequence features, and performing classification according to the feature vector to obtain a classification result of pipeline defects. According to the method, the acoustic signals are converted into different types of spectrograms, the original time sequence signals are combined, feature extraction, fusion and classification are carried out through a deep learning model, complementary information of the spectrograms and the time sequence signals is fully utilized, the limitation of a single recognition mode is overcome, and the accuracy and robustness of pipeline defect detection are improved.

The invention relates to the technical field of mass spectrum analysis, in particular to an analysis method and system for improving the spectrum quality of a mass spectrum. The method comprises the following steps: scanning a sample by using a mass spectrometer, collecting original mass spectrogram data, and preprocessing the original mass spectrogram data; based on the preprocessed original mass spectrum data, identifying potential peak positions through a multi-scale Gaussian fitting technology, and screening out candidate peaks by calculating the signal-to-noise ratio of each peak; fitting each screened candidate peak by using an asymmetric Gaussian function model, and describing the shape of the candidate peak; and constructing a mass offset correction model to correct the mass spectrum data, and performing super-resolution processing on the corrected mass spectrum data by using a convolutional neural network model. According to the design, peak fitting is carried out by adopting a multi-scale Gaussian fitting technology, a signal-to-noise ratio screening mechanism and an asymmetric Gaussian function model.

The invention relates to a radio frequency fingerprint identification method and system based on a convolution-attention mechanism and multi-packet reasoning, and belongs to the technical field of communication networks and artificial intelligence. Comprising the following steps: step 1, capturing an equipment transmission signal, and preprocessing the transmission signal to obtain a spectrogram; step 2, constructing a radio frequency fingerprint identification model, and inputting the obtained spectrogram into the radio frequency fingerprint identification model for training to obtain a trained radio frequency fingerprint identification model; and step 3, performing prediction by using the trained radio frequency fingerprint identification model to obtain a final prediction result. Aiming at the problems of high calculation complexity and redundant module interaction existing in a traditional model fusion mechanism, a module separation strategy is adopted, so that a convolution module and a Transform module are independently optimized, the calculation complexity is reduced, and the stability of the model is improved.

The invention discloses a full-automatic cable vibration high-precision measurement method based on a machine vision technology, and relates to the technical field of structural vibration measurement. The method comprises the following steps: installing and calibrating equipment, and ensuring that a cable plane coincides with an imaging plane; high-frame-rate video acquisition is carried out, and comparison verification is carried out in combination with an accelerometer; screening effective pixels based on the pixel intensity time sequence variance, and constructing a pixel intensity space-time matrix by using the time sequence of the pixels; sVD decomposition is adopted to extract an accumulated energy ratio gt; 90% of the main modes are reconstructed, and real signals are revealed; a plurality of inhaul cables are automatically positioned and coded through gradient features and Hough transform; sub-pixel-level vertical displacement is calculated by applying an optical flow method, and median filtering is combined to resist noise; fourier transform is carried out to obtain a spectrogram, and a fundamental frequency is identified; and calculating cable force based on a string pulling theorem. According to the method, remote, full-field and multi-target synchronous monitoring is realized, the displacement measurement precision reaches a sub-pixel level, and a high-precision and full-automatic solution is provided for health monitoring of structures such as bridges.

The invention discloses a multi-source federal cross-domain and source-domain enhanced millimeter wave action recognition method and system, and the method comprises the steps: generating a micro-Doppler spectrogram through a millimeter wave radar, and extracting the motion features of human body actions through a signal processing module; in the federal multi-source domain adaptation module, dynamically evaluating and fusing knowledge of a plurality of source domains by adopting a voting-based pseudo-tag method and a weighted knowledge aggregation mechanism, and optimizing the generalization ability of a target model; and through a generalization gap optimization method, the performance of the source domain model is improved, and the robustness of the system in different environments is ensured. Through combination of a federated learning framework and a multi-source domain adaptation technology, unsupervised learning under the condition that a target domain has no annotated data is realized, only a single set of millimeter wave equipment is needed, a millimeter wave communication protocol is compatible, and the method has the characteristics of privacy protection, unsupervised learning, multi-source knowledge fusion and strong generalization ability. The method is suitable for application scenes of smart home, health monitoring, man-machine interaction and the like, and has wide practical application value and research prospect.

The invention relates to the technical field of artificial intelligence, and discloses a Bluetooth communication intelligent speech translation method and system based on multi-mode enhancement, and the method comprises the steps: collecting a multi-channel audio signal through a built-in multi-microphone array of a Bluetooth device, carrying out the dynamic direction self-adaptive beam forming of the multi-channel audio signal, and carrying out the self-adaptive beam forming of the multi-channel audio signal; extracting a Mel spectrogram feature of the direction enhancement signal, identifying lip regions of a plurality of candidate speakers in each frame of real-time speaking video captured by a camera, performing time sequence convolution on the lip regions to obtain a lip movement time sequence embedded vector, calculating a correlation score with the Mel spectrogram feature, separating the direction enhancement signal, and obtaining a lip movement time sequence embedded vector; and performing text transcription and conversion on the high-confidence separation voice to obtain a translation language text, and sending the synthesized target translation voice to a preset mobile terminal through the Bluetooth device to obtain a target translation result. According to the method, the real-time performance and accuracy of speech translation are improved in a multi-person scene, far-field speech, noise interference and accent difference.

The invention relates to the related field of underwater acoustic signal processing and noise recognition, in particular to an underwater acoustic radiation noise recognition method, and provides an underwater acoustic radiation noise real-time recognition system based on combination of a convolutional neural network (CNN) and a long short-term memory network (LSTM). The system can perform efficient data preprocessing, feature extraction and optimization processing after collecting underwater acoustic signals in real time, and performs rapid identification of noise types through the trained mixed deep learning model. Specifically, the CNN can extract local features from a spectrogram, the LSTM is good at capturing time sequence information, complex underwater acoustic signals can be better understood and distinguished through combination of the CNN and the LSTM, and the recognition precision is remarkably improved. In addition, the system is further provided with an automatic updating mechanism, the model can be retrained regularly according to newly-collected data, it is ensured that the model is always in the optimal state, and the system adapts to the constantly-changing underwater environment.

Disclosed are systems and methods including software processes executed by a server that detect audio-based synthetic speech (“deepfakes”). Embodiments implement a machine-learning architecture having a diffusion model that generates purified features that are fed to a deepfake detection model. The machine-learning architecture includes input layers that convert an audio signal into a Gaussian or frequency space representation (e.g., log spectrogram) to extract a set of initial features indicative of spoofing or deepfake attacks. The diffusion model identifies adversarial noise on the audio signal in the initial features and generates purified features or clean version of the input audio signal. A deepfake detector includes a neural network architecture and classifier programmed and trained to generate a deepfake detection score and classify the audio signal as genuine or fraudulent using the purified features.

The invention discloses a target detection method and system based on a millimeter wave radar. The target detection method and system are used for realizing accurate detection, positioning and dynamic and static recognition of multiple targets under a complex background. According to the method, a distance-Doppler spectrogram is generated through the technical means of sliding window construction, spectral analysis, clutter suppression and the like, and candidate target points are detected by adopting an SO-CFAR algorithm. Then, determining a target position through high-resolution direction estimation and coordinate transformation, performing spatial clustering in combination with a density-based DBSCAN algorithm, and extracting a target geometric center and a bounding box; in the aspect of target tracking, Kalman filtering is used for predicting and updating the position and speed of the target, and a beam forming technology is used for enhancing a target signal, so that the target recognition stability is improved. And finally, the system performs robust dynamic and static state recognition on the target through a dynamic and static judgment module, so that high precision and robustness of the target detection process are ensured. The method can effectively cope with static background interference and dynamic target changes, and is suitable for target detection and tracking in a complex environment.

The invention relates to a power transmission line fault diagnosis method based on time-frequency multistage fusion and a bidirectional time sequence enhancement network, and belongs to the field of power system power transmission line fault diagnosis. The method comprises the following steps: firstly, constructing a time-frequency multi-level fusion and bidirectional time sequence enhancement network: converting input power transmission line partial discharge signal data into a time-frequency spectrogram; constructing a multi-scale time-frequency feature joint extraction module, and performing feature extraction on the time-frequency spectrogram by dual offset branches; constructing a dynamic weighted fusion module, and carrying out adaptive fusion on the extracted time-frequency features; a bidirectional time sequence information enhancement module is constructed, modeling is performed on the fusion features through a bidirectional long-short-term memory network, and feature expression of key discharge moments is enhanced by adopting an attention guiding mechanism; and finally, training the time-frequency multi-level fusion and bidirectional time sequence enhancement network by using the training set, and finally realizing fault type identification through a full connection layer and a Softmax classifier. According to the invention, the accuracy of power transmission line fault diagnosis and the model robustness are improved.

The invention relates to the technical field of sound quality enhancement, in particular to a sound quality improvement method and device of a Bluetooth loudspeaker box, equipment and a storage medium. The method comprises the following steps: collecting sound box environment acoustic signals, and carrying out peripheral audio time-frequency spectrum analysis and three-dimensional sound field fitting to construct a peripheral three-dimensional sound field model; based on the surrounding three-dimensional sound field model, environment noise source semantic deep analysis is carried out, and user real-time scene sound quality demand prediction is carried out, so that real-time scene sound quality demand features are generated; bluetooth sound box loudspeaker monitoring parameters are collected, and adaptive noise frequency suppression adjustment is carried out based on the surrounding three-dimensional sound field model rate, so that a loudspeaker dynamic noise suppression adjustment strategy is generated; and acquiring a Bluetooth audio output signal, and performing band frequency response characteristic optimization according to the real-time scene tone quality demand characteristics to obtain a frequency response tone quality optimized audio signal. According to changes of different environments and requirements, the output tone quality of the Bluetooth loudspeaker box is optimized in real time, and high-fidelity audio output is kept.

A method for monitoring a rotating machine in order to detect a fault in a bearing, the method including acquiring, from the rotating machine, a vibration signal measured by a vibration sensor; determining a first-order spectrogram by first-order cyclostationary analysis of the vibration signal using a delta transform and spectral standardisation; determining a second-order spectrogram by second-order cyclostationary analysis of the vibration signal using averaged cyclic coherence, a delta transform and spectral standardisation; and detecting a vibration signature of the fault in the bearing on the basis of the first-order spectrogram and the second-order spectrogram.

Disclosed are a multi-mode emotion recognition method, a system, an electronic device, and a storage medium. The method includes obtaining a spectrogram of a voice to be recognized and a corresponding text and inputting the spectrogram and the text into a multi-mode emotion recognition model to obtain an emotion recognition result output by the multi-mode emotion recognition model. The multi-mode emotion recognition model is trained based on a sample spectrogram, and a corresponding sample text, and a sample emotion recognition result, and is configured to extract a feature from the spectrogram and the text by a self-attention mechanism to obtain the voice features and the text feature, fuse the text feature and voice feature to obtain a multi-mode fusion feature, and make an emotion classification decision to obtain an emotion recognition result based on the text feature, the voice feature, and the multi-mode fusion feature.

The invention provides a vocal music training vowel pronunciation quality evaluation method based on auditory and visual spatial-temporal feature fusion, and the method comprises the steps: collecting vowel pronunciation audio signals and corresponding videos of a singer, and constructing a multi-modal data set; generating a fractional order Mel spectrogram for the audio signal through short-time fractional order Fourier transform of an adaptive order; extracting time sequence features and spatial features of the fractional order Mel spectrogram, and fusing the time sequence features and the spatial features through a gating mechanism to generate audio spatio-temporal features; face visual features in the video are extracted and fused with the audio spatio-temporal features through a cross attention mechanism, and the cross attention mechanism is integrated with a periodic modeling network; the fused features are input into a classifier, a dynamic weight multi-mode cosine loss function training model is adopted, the dynamic weight multi-mode cosine loss function dynamically adjusts the sample weight through a confusion matrix, and the weight is increased for the samples with classification errors based on the historical frequency mistaken division times of the samples; and outputting a pronunciation quality evaluation result.

The invention discloses a light-weight YOLOv8n model partial discharge type detection and classification method based on a PRPD spectrogram, and aims to solve the problems of low partial discharge type detection precision, high model complexity, large calculation amount, poor real-time performance and the like in the prior art. According to the method, PRPD spectrograms of four typical partial discharge types of tip, air gap, suspension and surface are collected, normalization, gray processing, data enhancement and expansion and other preprocessing are carried out, a LabelImg platform is utilized to carry out labeling, then a YOLOv8n model is subjected to lightweight improvement, ShuffleNet-V2 is adopted as a trunk feature extraction network, a CoordAttention mechanism is introduced to enhance the feature extraction capability, and the feature extraction efficiency is improved. An EIOU loss function is applied to optimize target frame regression precision; the partial discharge type detection and classification precision is improved, the model calculation complexity and parameter quantity are reduced, real-time detection is realized, and the method has important practical application value.