Online management method and system based on crew examination

Abstract

The invention relates to the technical field of sailor assessment management, and particularly discloses an online management method and system based on sailor assessment, and the method comprises the steps: collecting an original voice signal during sailor assessment, and a propeller rotating speed, a sea condition grade and a ship navigational speed at a corresponding moment; the original voice signals and the physical parameters are input into a physical constraint noise reduction network to serve as a regularization item embedding loss function, ocean background noise is separated, and pure sailor voice signals are output; inputting the pure sailor voice signals into a navigation term recognition network, calculating the similarity between voice features and IMO standard term prototype vectors, and outputting term accuracy scores and term confusion point information; based on the pure crew voice signal, the term accuracy score and the sea condition level, a voice signal-to-noise ratio, term accuracy and a sea condition influence factor are fused to calculate a communication capability index; generating a personalized performance evaluation report and pushing training courses; according to the invention, the problems of difficult voice noise reduction and disjunction between evaluation and training in a complex marine environment are solved.

Description

Technical Field

[0001] This invention relates to the field of seafarer assessment and management technology, specifically to a method and system for online seafarer assessment management. Background Technology

[0002] With the advancement of digital transformation in the shipping industry, some maritime authorities have introduced online examination systems and remote monitoring platforms, achieving paperless theoretical examinations and centralized control of examination room videos. Traditional crew assessment and management mainly adopted offline centralized examinations and manual evaluations, where examiners scored crew members' English listening and speaking abilities on-site in the examination room, or recorded audio and then manually reviewed by evaluators.

[0003] Existing online management methods for crew assessment face a fundamental contradiction when processing speech signals in real marine environments: the conflict between non-stationary noise interference caused by extreme sea states and the fidelity of speech features. Specifically, traditional noise reduction algorithms tend to damage the consonant features of terms when suppressing propeller cavitation noise and random spectrum noise from wind and waves. Furthermore, the evaluation model focuses only on the speech content itself and cannot quantify the physical impact of sea state levels on communication quality. Consequently, the output assessment scores cannot truly reflect the actual performance capabilities of crew members in harsh marine environments. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for online management of crew member assessments to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions: A method for online management of crew member assessment includes the following steps: S1: Collect the raw voice signals generated by the crew in the assessment scenario, and simultaneously obtain the ship's navigation physical parameters at the corresponding time. The physical parameters include propeller speed, sea state level and ship speed. S2: Input the original speech signal and physical parameters into the physical constraint denoising network. During the training phase, the physical constraint denoising network embeds the marine acoustic physics equation as a regularization term into the loss function. The physical parameters guide the network to separate the marine background noise that conforms to the law of sound wave propagation from the original speech signal and output a clean crew speech signal. S3: Input the clean crew voice signal into the nautical terminology recognition network. The nautical terminology recognition network is based on prototype learning. By calculating the similarity between the voice features and the preset IMO standard terminology prototype vector, it identifies the accuracy of terminology use in the crew voice, marks easily confused terminology pairs, and outputs terminology accuracy scores and terminology confusion information. S4: Based on the clean crew voice signal, terminology accuracy score and sea state level, the crew's communication ability index is calculated using a comprehensive evaluation algorithm. The comprehensive evaluation algorithm integrates voice signal-to-noise ratio, terminology accuracy and sea state influence factor. S5: Based on the communication ability index and terminology confusion information, generate a personalized performance evaluation report, and automatically match and push corresponding online training course content according to the skill gaps in the report.

[0006] As a further aspect of the present invention: S2 specifically includes: Perform a short-time Fourier transform on the original speech signal to obtain the time-spectrum feature map; The physical parameters are mapped to physical feature vectors, and the physical feature vectors are concatenated with the time-spectrum feature map along the channel dimension to generate a joint feature tensor. The joint feature tensor is input into the physical constraint denoising network. During the training process, the loss function of the physical constraint denoising network includes a physical constraint term constructed based on the ocean acoustic attenuation equation. The physical constraint term guides the network to separate the ocean background noise that matches the physical parameters from the joint feature tensor and outputs the ocean background noise. The pure crew voice signal is obtained by subtracting ocean background noise from the original voice signal.

[0007] As a further aspect of the present invention: the output ocean background noise specifically includes: Based on physical parameters, the theoretical attenuation coefficients of each frequency point at different propagation distances are calculated using the ocean acoustic attenuation equation, and the theoretical attenuation coefficients are arranged in order of frequency points to form a physical constraint weight matrix. The joint feature tensor is multiplied element-wise with the physical constraint weight matrix to obtain the physical weighted feature tensor. The physical weighted feature tensor is input into the physical constraint denoising network. During the training process, the loss function of the physical constraint denoising network includes a physical constraint term constructed based on the ocean acoustic attenuation equation. The physical constraint term is used to guide the network parameters to learn to extract noise components that match the physical parameters from the physical weighted feature tensor and output ocean background noise.

[0008] As a further aspect of the present invention: S3 specifically includes: The pure crew member speech signal is subjected to spectrogram conversion to obtain spectrogram feature map, and syllable boundary detection is performed on the spectrogram feature map to segment it into several candidate term segments; Each candidate term fragment is input into a feature extraction network to extract the corresponding term acoustic feature vector. The cosine similarity between the acoustic feature vector of each term and the standard term prototype vector in the preset IMO standard term prototype library is calculated, and the term corresponding to the standard term prototype vector with the highest cosine similarity is taken as the recognition result; the term corresponding to the standard term prototype vector with the second highest cosine similarity is extracted as the easily confused candidate term. The recognition result is compared with the standard answer. If they are inconsistent, the similarity difference between the acoustic feature vector of the term and the standard term prototype vector corresponding to the recognition result, as well as the standard term prototype vector corresponding to the easily confused candidate term, is calculated. When the similarity difference is less than the preset threshold, the recognition result and the easily confused candidate term are marked as an easily confused term pair, and the term accuracy score and the confusion point information of the term containing the easily confused term pair are output.

[0009] As a further aspect of the present invention: the step of inputting each candidate term fragment into a feature extraction network to extract the corresponding term acoustic feature vector specifically includes: The candidate term segments are segmented into multiple granularities to obtain both whole speech segments including complete terms and local speech segments including syllables. Mel frequency cepstral coefficients are extracted from the entire speech segment to obtain the overall feature vector, and linear predictive coding analysis is performed on the local speech segment to obtain the local feature vector reflecting the vocal tract resonance characteristics. Each frame in the local feature vector is arranged in chronological order to form a local temporal feature matrix; The overall feature vector and the local temporal feature matrix are concatenated and fused along the feature dimension, and the fused features are then subjected to an affine transformation to obtain the terminological acoustic feature vector.

[0010] As a further aspect of the present invention: S4 specifically includes: Speech activity detection is performed on the clean crew voice signal to extract effective speech segments, and the average signal-to-noise ratio of the effective speech segments is calculated as the speech signal-to-noise ratio. The accuracy of a term is calculated by taking the proportion of correct terms to all terms in the statistical term accuracy score. The term accuracy rate is then obtained by looking up the corresponding term weight coefficient from a pre-defined weight mapping table based on the numerical range of the term accuracy rate. The sea state level is input into the sea state influence function, which outputs the corresponding sea state influence coefficient based on the pre-defined correspondence between the sea state level and the influence coefficient. The communication capability index is obtained by weighting and fusing the voice signal-to-noise ratio, term weight coefficient, and sea state influence coefficient.

[0011] As a further aspect of the present invention: the output process of the sea state influence coefficient is as follows: Obtain the wave spectrum peak period and significant wave height corresponding to the sea state level, and calculate the wave energy density based on the wave spectrum peak period and the significant wave height; The wave energy density is input into the ship motion response function. The ship motion response function performs a weighted integral on the wave energy density according to the amplitude-frequency response operator corresponding to the ship type to obtain the ship heave motion amplitude. The amplitude of the ship's heave motion is input into the voice acquisition perturbation function. The voice acquisition perturbation function calculates the acceleration perturbation value generated at the microphone by the amplitude of the ship's heave motion based on the lever arm length between the microphone installation position and the ship's center of gravity. The acceleration disturbance value is input into the influence coefficient mapping network, which outputs the corresponding sea state influence coefficient based on the pre-calibrated correspondence between the acceleration disturbance value and the signal attenuation.

[0012] As a further aspect of the present invention: S5 specifically includes: Based on the information on confusing terms, several knowledge point nodes associated with the confusing terms are extracted from the pre-set nautical terminology knowledge graph, and these knowledge point nodes are sorted according to their degree of association to obtain a knowledge point sequence. The communication capability index is compared with multiple preset capability level thresholds to determine the current capability level of the crew member, and the corresponding difficulty of the knowledge points to be strengthened is selected from the knowledge point sequence according to the capability level. The knowledge points to be strengthened are input into the training content matching library. Each training content in the training content matching library is pre-labeled with knowledge point tags and difficulty tags. By matching the knowledge points to be strengthened with the knowledge point tags, the training content that meets the difficulty tags is selected as candidate training content. The learning urgency coefficient is calculated based on the communication capability index, and the candidate training content is prioritized according to the learning urgency coefficient. The training content with the highest priority is selected to generate a push instruction, which is then sent to the crew terminal to present the corresponding online training course.

[0013] As a further aspect of the present invention: the step of matching the knowledge points to be reinforced with knowledge point tags to select training content that simultaneously meets the difficulty tags as candidate training content specifically includes: The knowledge points to be strengthened are input into the nautical terminology association network. Each node in the nautical terminology association network corresponds to a terminology knowledge point. The edges between nodes are pre-weighted according to the co-occurrence frequency of the terms in nautical operations. By traversing the nodes corresponding to the knowledge points to be strengthened, the adjacent nodes connected to the nodes and whose edge weights are greater than a preset threshold are extracted to form an extended knowledge point set. Each extended knowledge point in the extended knowledge point set is compared one by one with the knowledge point tags in the training content matching library. Training content whose knowledge point tags include extended knowledge points is selected to obtain the initial set of training content. Read the pre-marked difficulty tags of each training content in the initial selection training content set, compare the difficulty tags with the corresponding ability level thresholds of the knowledge points to be strengthened, and remove the training content whose difficulty tags are lower than the ability level thresholds to obtain candidate training content.

[0014] An online management system for crew assessment includes: The data acquisition module is used to collect the raw voice signals generated by the crew in the assessment scenario and simultaneously acquire the ship's navigation physical parameters at the corresponding time. The physical parameters include propeller speed, sea state level and ship speed. The noise reduction module inputs the original speech signal and physical parameters into the physical constraint noise reduction network. During the training phase, the physical constraint noise reduction network embeds the marine acoustic physics equation as a regularization term into the loss function and uses the physical parameters to guide the network to separate the marine background noise that conforms to the law of sound wave propagation from the original speech signal and output a clean crew speech signal. The terminology recognition module inputs the clean crew voice signal into the maritime terminology recognition network. The maritime terminology recognition network is based on prototype learning. By calculating the similarity between the voice features and the preset IMO standard term prototype vectors, it identifies the accuracy of terminology use in the crew's voice, marks easily confused term pairs, and outputs terminology accuracy scores and terminology confusion information. The comprehensive evaluation module calculates the crew's communication ability index based on the clean crew voice signal, terminology accuracy score and sea state level. The comprehensive evaluation algorithm integrates voice signal-to-noise ratio, terminology accuracy and sea state influence factor. The reporting and push module generates personalized performance evaluation reports based on communication ability index and terminology confusion information, and automatically matches and pushes corresponding online training course content based on the skill gaps in the report.

[0015] The beneficial effects of this invention are: (1) This invention introduces marine acoustic physics equations to physically constrain the noise reduction process and uses synchronously collected physical parameters such as propeller speed and sea state level to guide the network to separate marine background noise that conforms to the law of sound wave propagation. This solves the problem that traditional noise reduction algorithms are prone to damaging speech features under extreme sea conditions, significantly improves the purity of crew voice signals in complex marine environments, provides a high-fidelity data foundation for subsequent terminology recognition and communication capability assessment, and ensures the accuracy and reliability of the assessment results.

[0016] (2) This invention constructs a maritime terminology recognition network based on prototype learning. It achieves accurate terminology recognition by calculating the similarity between speech features and IMO standard terminology prototype vectors, and automatically marks easily confused terminology pairs. At the same time, it combines communication capability index and terminology confusion point information, expands knowledge points through maritime terminology knowledge graph, accurately matches and pushes personalized training courses based on learning urgency coefficient, realizes closed-loop management from assessment to skills improvement, and effectively improves the pertinence and training efficiency of crew members' on-the-job learning. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system block diagram of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 As shown, this invention is an online management method for crew member assessment, comprising the following steps: S1: Collect the raw voice signals generated by the crew in the assessment scenario, and simultaneously obtain the ship's navigation physical parameters at the corresponding time. The physical parameters include propeller speed, sea state level and ship speed. S2: Input the original speech signal and physical parameters into the physical constraint denoising network. During the training phase, the physical constraint denoising network embeds the marine acoustic physics equation as a regularization term into the loss function. The physical parameters guide the network to separate the marine background noise that conforms to the law of sound wave propagation from the original speech signal and output a clean crew speech signal. S3: Input the clean crew voice signal into the nautical terminology recognition network. The nautical terminology recognition network is based on prototype learning. By calculating the similarity between the voice features and the preset IMO standard terminology prototype vector, it identifies the accuracy of terminology use in the crew voice, marks easily confused terminology pairs, and outputs terminology accuracy scores and terminology confusion information. S4: Based on the clean crew voice signal, terminology accuracy score and sea state level, the crew's communication ability index is calculated using a comprehensive evaluation algorithm. The comprehensive evaluation algorithm integrates voice signal-to-noise ratio, terminology accuracy and sea state influence factor. S5: Based on the communication ability index and terminology confusion information, generate a personalized performance evaluation report, and automatically match and push corresponding online training course content according to the skill gaps in the report.

[0021] In S1, the raw voice signals generated by the crew during the assessment scenario are collected, and the ship's navigation physical parameters at the corresponding time are acquired simultaneously. These physical parameters include propeller speed, sea state, and ship speed, specifically including: First, a miniature microphone array is deployed near the steering position on the ship's bridge. The array's pickup range covers the crew's usual position when conducting VHF communications. When the crew begins answering communication scenario questions randomly selected by the assessment system, the microphone array captures the crew's continuous speech at a sampling rate of 16kHz, generates the raw speech signal, and stores it in real time as a WAV audio file.

[0022] Simultaneously, propeller speed data is obtained through the ship's existing engine room monitoring and alarm system. Specifically, the output value of the stern shaft speed sensor is read once per second via the data interface of the engine room control system to obtain the real-time propeller speed in revolutions per minute. The ship's speed is obtained through the ship's GPS receiver, reading the pulse signal of the water speed and converting it to obtain the instantaneous ship speed in knots.

[0023] In addition, the sea state level of the current sea area is obtained through the wave monitoring module integrated in the ship's weather fax receiver or shipborne navigation data recorder. This sea state level is automatically generated based on wave height observations and the international sea state scale. Specifically, the significant wave height value output by the wave meter is input into the sea state level determination unit, which outputs the corresponding sea state level value for the Beaufort scale.

[0024] To ensure data consistency, all data was assigned a unified timestamp during acquisition. The ship's time server obtained standard time through the BeiDou Navigation Satellite System and synchronously triggered the microphone array, engine room monitoring system, and GPS receiver using pulse signals, ensuring strict alignment between the start time of the raw voice signal and the acquisition time of each physical parameter. After acquisition, the raw voice signal, propeller speed, sea state, and ship speed were encapsulated into data packets and temporarily stored in the industrial control computer on the bridge, awaiting uploading to the shore-based server via the shipborne satellite communication system.

[0025] In S2, the original speech signal and physical parameters are input into a physically constrained noise reduction network. During the training phase, the network embeds the ocean acoustic physics equations as regularization terms into the loss function. The physical parameters guide the network to separate ocean background noise that conforms to the laws of sound wave propagation from the original speech signal, outputting a clean crew speech signal. Specifically, this includes: First, a short-time Fourier transform is performed on the acquired raw speech signal. Specifically, the raw speech signal is divided into frames with a length of 25 milliseconds and a frame shift of 10 milliseconds, and a Hanning window is applied to each frame to reduce spectral leakage. A fast Fourier transform is then performed on each windowed frame to convert the time-domain signal into a frequency-domain representation, resulting in a two-dimensional complex matrix composed of frequency components and time frames. Taking the modulus of this complex matrix yields a time-spectral feature map with frequency as the vertical axis and time as the horizontal axis.

[0026] Secondly, the three physical parameters of propeller speed, sea state, and ship speed, which were collected synchronously, were numerically normalized. Specifically, the normalization method was as follows: for propeller speed, the measured value was divided by 3000 revolutions per minute (RPM) as the baseline, yielding a normalized speed value between 0 and 1; for sea state, the highest level (level 9) was used as the baseline, and the measured level was divided by 9 to obtain a normalized sea state value; for ship speed, the maximum speed (30 knots) was used as the baseline, and the measured speed was divided by 30 to obtain a normalized speed value. These three normalized values ​​were arranged sequentially to form a three-dimensional physical feature vector. This physical feature vector was then copied and expanded along the channel dimension to match the size of the time-spectrum feature map. Finally, the expanded physical feature vector and the time-spectrum feature map were concatenated along the channel axis to generate a joint feature tensor containing both audio domain information and physical parameter information.

[0027] Next, the physical constraint weight matrix is ​​calculated based on the physical parameters. Based on the spherical expansion attenuation law in marine acoustics, the sound pressure level decreases in seawater propagation according to the inverse square law of distance as the propagation distance increases. First, the propagation distance between the cavitation noise source and the microphone is estimated based on the propagator speed. Specifically, when the propagator speed is greater than 200 revolutions per minute, cavitation noise is considered significant, and the propagation distance is taken as the straight-line distance from the center of the propagator disk to the bridge microphone, a fixed value determined during the ship design phase. When the sea state is greater than level 5, the wind and wave noise source is considered to mainly originate from the sea surface, and the propagation distance is taken as the combined vertical and horizontal distance from the wave-breaking zone to the bridge microphone. Then, for each frequency point in the time-spectrum characteristic graph, the theoretical attenuation coefficient corresponding to that frequency point is calculated based on the frequency value and the estimated propagation distance, using the empirical absorption and attenuation coefficient of sound waves in seawater. The theoretical attenuation coefficients of all frequencies are arranged in ascending order of frequency to form a one-dimensional physical constraint weight matrix.

[0028] Then, the joint feature tensor is multiplied element-wise with the physical constraint weight matrix. Specifically, the physical constraint weight matrix is ​​copied along the time dimension to match the time frame of the joint feature tensor, forming a weight tensor of the same size as the joint feature tensor. This weight tensor is then multiplied by each value at the corresponding position in the joint feature tensor to obtain the physically weighted feature tensor. In this physically weighted feature tensor, frequency components that conform to the attenuation law of sound wave propagation are preserved, while frequency components that deviate from this law are suppressed.

[0029] Finally, the physically weighted feature tensor is input into the physically constrained denoising network. This network adopts an encoder-decoder structure, consisting of multiple stacked convolutional and deconvolutional layers. During the network training phase, the loss function consists of a weighted sum of two parts: the first part is the reconstruction loss, which calculates the mean square error between the network output ocean background noise and the pre-labeled real noise; the second part is the physical constraint term, which is constructed by performing a short-time Fourier transform on the network output ocean background noise to obtain its spectrum, multiplying this spectrum by the physical constraint weight matrix, and calculating the energy sum, which is the value of the physical constraint term. The role of the physical constraint term is to guide the network parameters to prioritize learning and outputting noise components whose spectral distribution is consistent with the physical attenuation law during the update process. After training, the physically constrained denoising network, upon receiving the physically weighted feature tensor, extracts and outputs ocean background noise that matches the current physical parameters through layer-by-layer convolution calculation. Then, the time-domain waveform of the ocean background noise is subtracted from the time-domain waveform of the original speech signal point by point to obtain the clean crew speech signal.

[0030] In S3, the clean crew voice signal is input into the maritime terminology recognition network. Based on prototype learning, this network identifies the accuracy of terminology usage in the crew's speech by calculating the similarity between speech features and preset IMO standard terminology prototype vectors. It also identifies easily confused terminology pairs and outputs a terminology accuracy score and terminology confusion information, specifically including: First, spectrogram transformation is performed on the purified crew voice signal obtained after noise reduction. Specifically, the purified crew voice signal is divided into frames with a length of 20 milliseconds and a frame shift of 10 milliseconds. A Hamming window is applied to each frame to reduce spectral leakage, and then a Fast Fourier Transform is performed to obtain the spectral amplitude of each frame. The spectral amplitudes of all frames are arranged in chronological order to form a two-dimensional grayscale image with frequency as the vertical axis and time as the horizontal axis, i.e., the spectrogram feature map. Syllable boundary detection is performed on this spectrogram feature map: the short-time energy and short-time zero-crossing rate of each frame are calculated, and the energy threshold is set to 30% of the maximum energy, and the zero-crossing rate threshold is set to 50 times per second. When the short-time energy of three consecutive frames is lower than the energy threshold and the short-time zero-crossing rate is higher than the zero-crossing rate threshold, this is determined to be the end of a syllable; when the short-time energy of the subsequent three consecutive frames is higher than the energy threshold and the short-time zero-crossing rate is lower than the zero-crossing rate threshold, this is determined to be the beginning of the next syllable. Based on the detected syllable start and end points, the spectrogram feature map is divided into several candidate term segments, each segment corresponding to an independent term pronunciation unit.

[0031] Secondly, feature extraction is performed on each candidate term segment to obtain the corresponding term acoustic feature vector. This involves the following four sub-steps: The first sub-step involves multi-granular segmentation of the candidate term segments. The candidate term segments are retained as a whole, serving as a complete speech segment containing the complete term. Simultaneously, a sliding window with a fixed window length of 30 milliseconds and a window shift of 10 milliseconds is used to continuously extract the candidate term segments, resulting in multiple local speech segments containing syllables.

[0032] The second sub-step involves extracting Mel-frequency cepstral coefficients from the entire speech segment. The specific calculation process is as follows: First, the entire speech segment is pre-emphasized with a pre-emphasis coefficient of 0.97 to enhance high-frequency components. Then, frame-by-frame windowing is applied, with each frame being 25 milliseconds long and shifted by 10 milliseconds, using a Hamming window. A Fast Fourier Transform is performed on each frame to obtain the spectrum. The spectral energy is then filtered through a 24-channel Mel-filter bank to obtain the output energy of each filter. The logarithm of the output energy of the 24 filters is taken, followed by a Discrete Cosine Transform. The 2nd to 13th coefficients, a total of 12 dimensions, are used to construct the Mel-frequency cepstral coefficients for each frame. The 12-dimensional coefficients of all frames are arranged in chronological order to obtain the overall feature vector, whose dimension is the frame number multiplied by 12.

[0033] The third sub-step involves performing linear predictive coding analysis on local speech segments. For each local speech segment, a 10th-order linear predictive analysis is used. The calculation process is as follows: an autocorrelation matrix is ​​constructed, and the linear prediction coefficients are solved using the Durbin recursive algorithm to obtain 10 prediction coefficients as the linear predictive coding coefficients for that local speech segment; simultaneously, the short-time energy of the segment is calculated, and these 10 coefficients are concatenated with the short-time energy to obtain an 11-dimensional local feature vector. This process is repeated for all local speech segments to obtain a series of 11-dimensional local feature vectors.

[0034] The fourth sub-step involves arranging all the local feature vectors obtained in the third sub-step in chronological order to form a local temporal feature matrix with a dimension equal to the number of local segments multiplied by 11. This local temporal feature matrix is ​​then interpolated along the temporal dimension to ensure its frame count matches the frame count of the overall feature vector. The adjusted local temporal feature matrix is ​​then concatenated with the overall feature vector along the feature dimension to form a joint feature matrix with a dimension equal to the frame count multiplied by 23. This joint feature matrix is ​​then input into a feature transformation layer consisting of two fully connected layers. The first fully connected layer maps the 23-dimensional input to 128 dimensions, and the second layer maps the 128-dimensional input to 64 dimensions. After processing by a linear rectified activation function, a 64-dimensional terminological acoustic feature vector is finally output.

[0035] Next, terminology recognition and confusion labeling are performed. A pre-defined IMO standard terminology prototype library is used, containing 200 commonly used terms in standard maritime communication. Each term corresponds to a 64-dimensional standard term prototype vector. These prototype vectors are obtained by collecting standard pronunciation samples from 100 experienced seafarers, extracting 64-dimensional vectors from each sample using the aforementioned feature extraction method, and then averaging the vectors of all samples for the same term. For each candidate term segment, the cosine similarity between its 64-dimensional acoustic feature vector and each standard term prototype vector in the library is calculated. The cosine similarity is calculated by dividing the dot product of the term's acoustic feature vector and the standard term prototype vector by the product of the magnitude of the term's acoustic feature vector and the magnitude of the standard term prototype vector. The 200 similarity values ​​are sorted from highest to lowest, and the standard term corresponding to the highest similarity value is taken as the recognition result, while the standard term corresponding to the second highest similarity value is taken as the easily confused candidate term.

[0036] Finally, accuracy is assessed and confusion points are output. The recognition result is compared with the pre-set standard answer for the question. If they match, the term is correctly recognized, and the term accuracy score is increased by 1 point. If they do not match, the similarity value between the acoustic feature vector of the term and the standard term prototype vector corresponding to the recognition result is calculated, and then the similarity value between the acoustic feature vector of the term and the standard term prototype vector corresponding to the easily confused candidate term is subtracted to obtain the similarity difference. A preset threshold of 0.15 is set. When the similarity difference is less than 0.15, the two terms are considered to be highly similar in acoustic features, and the recognition result and the easily confused candidate term are marked as a pair of easily confused terms. After traversing all candidate term segments, the percentage of correctly recognized terms out of all assessed terms is counted as the term accuracy score; at the same time, all marked easily confused term pairs are compiled into term confusion point information and output together.

[0037] In S4, based on clean crew voice signals, terminology accuracy scores, and sea state levels, a comprehensive evaluation algorithm is used to calculate the crew's communication capability index. This algorithm integrates voice signal-to-noise ratio, terminology accuracy, and sea state influencing factors, specifically including: First, speech activity detection is performed on the clean crew voice signal to extract valid speech segments. The specific detection process is as follows: the clean crew voice signal is divided into frames with a length of 20 milliseconds and a frame shift of 10 milliseconds. The short-time energy and short-time zero-crossing rate of each frame are calculated. An energy threshold is set to 30% of the average energy of all voice frames; a zero-crossing rate threshold is set to 25 times per second. When the short-time energy of 5 consecutive frames is greater than the energy threshold and the short-time zero-crossing rate of 5 consecutive frames is less than the zero-crossing rate threshold, this is determined to be the start point of the speech; when the short-time energy of 5 consecutive frames is less than the energy threshold and the short-time zero-crossing rate of 5 consecutive frames is greater than the zero-crossing rate threshold, this is determined to be the end point of the speech. Based on the detected start and end points, valid speech segments are extracted from the clean crew voice signal. To calculate the average signal-to-noise ratio of the effective speech segment: take the average energy of the speech frames in the effective speech segment as the signal energy, and take the average energy of the silent segment (i.e., multiple consecutive frames before the start of the speech or after the end of the speech) as the noise energy. Calculate the logarithm of the ratio of signal energy to noise energy with base 10, and then multiply by 10 to obtain the speech signal-to-noise ratio in decibels.

[0038] Secondly, the accuracy score of the terminology is calculated, and the terminology weight coefficient is retrieved. The terminology accuracy score records the number of terms correctly identified by the crew member during the assessment. Assuming the assessment includes 20 terms, dividing the number of correct terms by 20 yields a terminology accuracy rate between 0 and 1. Based on the numerical range of the terminology accuracy rate, the corresponding terminology weight coefficient is retrieved from a pre-defined weight mapping table. This weight mapping table is constructed as follows: when the terminology accuracy rate is greater than or equal to 0.9, the corresponding terminology weight coefficient is 1.2; when the terminology accuracy rate is greater than or equal to 0.7 and less than 0.9, the corresponding terminology weight coefficient is 1.0; when the terminology accuracy rate is greater than or equal to 0.5 and less than 0.7, the corresponding terminology weight coefficient is 0.8; and when the terminology accuracy rate is less than 0.5, the corresponding terminology weight coefficient is 0.6. According to the above correspondence, the corresponding terminology weight coefficient is read from the weight mapping table based on the currently calculated terminology accuracy rate.

[0039] Then, the sea state influence coefficient is calculated, which includes the following four sub-steps: The first sub-step involves obtaining the wave parameters corresponding to the sea state level and calculating the wave energy density. Sea state levels are classified according to the internationally accepted sea state scale, ranging from 0 to 9. For a given sea state level, the corresponding wave spectral peak period and significant wave height are read from a preset wave parameter lookup table. For example, for sea state level 5, the wave spectral peak period is taken as 8.5 seconds, and the significant wave height as 3.5 meters. The wave energy density is calculated as follows: multiply the gravitational acceleration constant of 9.8 meters per square second by the square of the significant wave height, then divide by 16. The result is the wave energy density, expressed in square meters.

[0040] The second sub-step involves inputting the wave energy density into the ship's motion response function to calculate the ship's heave amplitude. For specific ship types, the amplitude-frequency response operator is pre-obtained through ship seakeeping experiments or hydrodynamic calculations. This operator is a function with wave frequency as the independent variable, representing the ship's heave response amplitude under wave action at different frequencies. For the current wave spectrum peak period, it is first converted to a circular frequency by dividing the wave spectrum peak period by twice pi. Based on this circular frequency value, the corresponding response amplitude is read from the amplitude-frequency response operator. The wave energy density is multiplied by this response amplitude, and the result is taken as the ship's heave amplitude, in meters.

[0041] The third sub-step involves inputting the ship's heave amplitude into the voice acquisition perturbation function to calculate the acceleration perturbation value at the microphone. First, the lever arm length between the microphone's installation location and the ship's center of gravity is obtained. This lever arm length is determined by the ship's design drawings. For example, if the microphone is installed on top of the bridge, at a horizontal distance of 15 meters and a vertical distance of 8 meters from the ship's center of gravity, the combined lever arm length is the square root of the sum of the squares of the horizontal and vertical distances, resulting in 17 meters. The voice acquisition perturbation function is calculated as follows: multiply the ship's heave amplitude by the square of twice pi, then divide by the square of the wave peak period to obtain the ship's heave acceleration amplitude; multiply this acceleration amplitude by the ratio of the lever arm length to the distance from the ship's center of gravity to the bow to obtain the acceleration perturbation value at the microphone, in meters per square second.

[0042] The fourth sub-step involves inputting the acceleration disturbance value into the influence coefficient mapping network and outputting the sea state influence coefficient. This influence coefficient mapping network is a pre-built lookup table, calibrated experimentally: In an experimental environment, a microphone is fixed to a vibration table, vibrations of different acceleration values ​​are applied, and standard test audio is played simultaneously. The energy attenuation ratio of the audio signal acquired under vibration conditions relative to the same audio signal acquired under no vibration conditions is measured. The acceleration disturbance value is calibrated from 0.1 m / s² to 2.0 m / s², with a step size of 0.1, to determine the corresponding signal attenuation ratio. This attenuation ratio is the sea state influence coefficient, ranging from 0.8 to 1.0. Based on the currently calculated acceleration disturbance value, the sea state influence coefficient corresponding to the closest acceleration disturbance value is found in the lookup table and output; if the acceleration disturbance value exceeds the calibration range, the sea state influence coefficient corresponding to the boundary value is used.

[0043] Finally, a weighted fusion operation is performed to obtain the communication capability index. This involves multiplying and fusing the speech signal-to-noise ratio (SNR), the terminology weight coefficient, and the sea state influence coefficient. Specifically, the SNR value is multiplied by the terminology weight coefficient, and the resulting product is then multiplied by the sea state influence coefficient. The final value is the crew member's communication capability index. This communication capability index is a dimensionless value used to comprehensively reflect the actual communication level of the crew member under specific sea state conditions.

[0044] In S5, a personalized performance evaluation report is generated based on the communication capability index and terminology confusion information. The report also automatically matches and pushes corresponding online training courses to address skill gaps identified in the report. These courses include: First, based on the terminology confusion information output in step S3, knowledge point nodes associated with the confused terms are extracted from a pre-defined maritime terminology knowledge graph. This maritime terminology knowledge graph is constructed as follows: each term in the IMO standard maritime terminology database is used as a node, and the edges between nodes represent the semantic associations of the terms in real maritime operations. The specific association degree is calculated based on large-scale ship logs, VHF communication records, and accident report texts, statistically analyzing the frequency of co-occurrence of each pair of terms in the same document or the same continuous communication. The co-occurrence frequency is normalized; that is, for each term, the number of times it co-occurs with all other terms is divided by the total number of times the term appears, resulting in an association degree value between 0 and 1, which is used as the edge weight. When terminology confusion information is received, each confused term involved is extracted, the corresponding node is located in the knowledge graph, and the neighboring nodes with a weight greater than a pre-defined threshold of 0.3 among all edges connected to that node are identified. These neighboring nodes are then used as associated knowledge points. All associated knowledge points are sorted from high to low according to their association degree with the original confused term, resulting in a knowledge point sequence.

[0045] Next, the communication capability index calculated in step S4 is compared with several preset capability level thresholds to determine the crew member's current capability level. The preset capability level thresholds are divided into three levels: when the communication capability index is greater than or equal to 80, it is considered advanced; when the communication capability index is greater than or equal to 60 and less than 80, it is considered intermediate; and when the communication capability index is less than 60, it is considered basic. Based on the determined capability level, knowledge points of corresponding difficulty are selected from the above knowledge point sequence as knowledge points to be strengthened. The specific selection rules are as follows: if the crew member is at the basic level, the first 5 knowledge points corresponding to basic terms (i.e., nodes with low in-degree in the knowledge graph that belong to core basic vocabulary) in the knowledge point sequence are selected; if at the intermediate level, the first 8 knowledge points of medium difficulty (i.e., between basic and advanced) in the sequence are selected; and if at the advanced level, the first 10 knowledge points of high professionalism (i.e., related to specific equipment and emergency operations) in the sequence are selected.

[0046] Next, the knowledge points to be reinforced are input into the training content matching database for matching. This database pre-stores multiple online training courses, each labeled with knowledge point tags and difficulty tags. The knowledge point tags are the set of nautical terms covered by the course, and the difficulty tags are divided into three levels: beginner, intermediate, and advanced, corresponding to the crew's competency level. The matching process specifically includes the following sub-steps: The first sub-step involves inputting the knowledge points to be reinforced into the nautical terminology association network. This network has the same structure as the knowledge graph described above, with each node corresponding to a terminology knowledge point. The edge weights between nodes are pre-assigned based on the co-occurrence frequency of the terms in nautical operations, with weight values ​​ranging from 0 to 1. An expansion threshold of 0.5 is set. For each node corresponding to the knowledge point to be reinforced, all directly connected neighboring nodes with edge weights greater than 0.5 are extracted. These neighboring nodes are then merged with the original knowledge point to be reinforced, forming an expanded knowledge point set. This set includes both the core terms that the crew currently needs to reinforce and expanded terms that are semantically and operationally closely related, ensuring the comprehensiveness of the training content.

[0047] The second sub-step involves comparing each extended knowledge point in the extended knowledge point set with the knowledge point tags in the training content matching library one by one. The comparison method is as follows: check if the training content's knowledge point tag set contains the current extended knowledge point; if it does, then include the training content in the initial selection training content set. After traversing all extended knowledge points, a preliminary selection training content set containing all matching courses is obtained.

[0048] The third sub-step involves reading the pre-marked difficulty tags for each training item in the initial selection set, and simultaneously obtaining the corresponding ability level thresholds for the knowledge points to be strengthened (i.e., the difficulty requirements corresponding to the crew member's current ability level). The difficulty tag for each training item is compared with the ability level threshold: if the difficulty tag is lower than the threshold (e.g., the crew member is intermediate level, but the course difficulty is beginner level), the course is too simple and cannot effectively improve the crew member's abilities, and is therefore eliminated; if the difficulty tag is equal to or higher than the threshold, the course is retained. After this screening, the final set of candidate training items is obtained.

[0049] Next, a learning urgency coefficient is calculated based on the communication capability index, which is used to prioritize candidate training content. The formula for calculating the learning urgency coefficient is as follows: ; in, This represents the learning urgency coefficient, with a value ranging from 0 to 1. The higher the value, the higher the training urgency. The communication capability index calculated in step S4 has a value range of 0 to 100. The number of easily confused term pairs marked in the term confusion point information; The total number of terms involved in the assessment is used to normalize the logarithm of confusing terms to between 0 and 1. and The weighting coefficients, 0.6 and 0.4 respectively, are used to balance the impact of overall communication competence and the degree of confusion regarding specific terms on urgency. This formula allows for a quantitative assessment of the extent of crew members' communication skill deficiencies and the density of confusing terminology, thereby determining training priorities.

[0050] Finally, according to the learning urgency coefficient The candidate training content is sorted from highest to lowest quality, and the top-ranked content is selected as the optimal content to push. A push instruction containing an identifier for this training content is generated and sent to the crew's terminal via the shipboard satellite communication network. Upon receiving the instruction, the crew's terminal automatically displays a reminder and loads the corresponding online training course page for targeted learning. Simultaneously, the push result and subsequent learning progress will be recorded for competency tracking and evaluation in the next assessment.

[0051] Please see Figure 2 As shown, an online management system for crew assessment includes: The data acquisition module is used to collect the raw voice signals generated by the crew in the assessment scenario and simultaneously acquire the ship's navigation physical parameters at the corresponding time. The physical parameters include propeller speed, sea state level and ship speed. The noise reduction module inputs the original speech signal and physical parameters into the physical constraint noise reduction network. During the training phase, the physical constraint noise reduction network embeds the marine acoustic physics equation as a regularization term into the loss function and uses the physical parameters to guide the network to separate the marine background noise that conforms to the law of sound wave propagation from the original speech signal and output a clean crew speech signal. The terminology recognition module inputs the clean crew voice signal into the maritime terminology recognition network. The maritime terminology recognition network is based on prototype learning. By calculating the similarity between the voice features and the preset IMO standard term prototype vectors, it identifies the accuracy of terminology use in the crew's voice, marks easily confused term pairs, and outputs terminology accuracy scores and terminology confusion information. The comprehensive evaluation module calculates the crew's communication ability index based on the clean crew voice signal, terminology accuracy score and sea state level. The comprehensive evaluation algorithm integrates voice signal-to-noise ratio, terminology accuracy and sea state influence factor. The reporting and push module generates personalized performance evaluation reports based on communication ability index and terminology confusion information, and automatically matches and pushes corresponding online training course content based on the skill gaps in the report.

[0052] The working principle of this invention is as follows: Raw voice signals generated by crew members during assessment scenarios are collected, and corresponding ship navigation physical parameters, including propeller speed, sea state, and ship speed, are acquired simultaneously. The raw voice signals and physical parameters are input into a physically constrained noise reduction network. During the training phase, this network embeds the ocean acoustic physics equation as a regularization term into the loss function. The physical parameters guide the network to separate ocean background noise conforming to the sound wave propagation laws from the raw voice signals, outputting a clean crew voice signal. This clean crew voice signal is then input into a nautical terminology recognition network. This network is based on prototype learning and uses computation... The similarity between speech features and preset IMO standard terminology prototype vectors is calculated to identify the accuracy of terminology use in crew members' speech, and easily confused terminology pairs are marked. The terminology accuracy score and terminology confusion point information are output. Based on the clean crew member speech signal, terminology accuracy score and sea state level, a comprehensive evaluation algorithm is used to calculate the crew member's communication ability index. The comprehensive evaluation algorithm integrates speech signal-to-noise ratio, terminology accuracy rate and sea state influence factor. Based on the communication ability index and terminology confusion point information, a personalized performance evaluation report is generated, and corresponding online training course content is automatically matched and pushed according to the skill gaps in the report.

[0053] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for online management of crew member assessment, characterized in that, Includes the following steps: S1: Collect the raw voice signals generated by the crew in the assessment scenario, and simultaneously obtain the ship's navigation physical parameters at the corresponding time. The physical parameters include propeller speed, sea state level and ship speed. S2: Input the original speech signal and physical parameters into the physical constraint denoising network. During the training phase, the physical constraint denoising network embeds the marine acoustic physics equation as a regularization term into the loss function. The physical parameters guide the network to separate the marine background noise that conforms to the law of sound wave propagation from the original speech signal and output a clean crew speech signal. S3: Input the clean crew voice signal into the nautical terminology recognition network. The nautical terminology recognition network is based on prototype learning. By calculating the similarity between the voice features and the preset IMO standard terminology prototype vector, it identifies the accuracy of terminology use in the crew voice, marks easily confused terminology pairs, and outputs terminology accuracy scores and terminology confusion information. S4: Based on the clean crew voice signal, terminology accuracy score and sea state level, the crew's communication ability index is calculated using a comprehensive evaluation algorithm. The comprehensive evaluation algorithm integrates voice signal-to-noise ratio, terminology accuracy and sea state influence factor. S5: Based on the communication ability index and terminology confusion information, generate a personalized performance evaluation report, and automatically match and push corresponding online training course content according to the skill gaps in the report.

2. The method for online management of crew member assessment according to claim 1, characterized in that, S2 specifically includes: Perform a short-time Fourier transform on the original speech signal to obtain the time-spectrum feature map; The physical parameters are mapped to physical feature vectors, and the physical feature vectors are concatenated with the time-spectrum feature map along the channel dimension to generate a joint feature tensor. The joint feature tensor is input into the physical constraint denoising network. During the training process, the loss function of the physical constraint denoising network includes a physical constraint term constructed based on the ocean acoustic attenuation equation. The physical constraint term guides the network to separate the ocean background noise that matches the physical parameters from the joint feature tensor and outputs the ocean background noise. The pure crew voice signal is obtained by subtracting ocean background noise from the original voice signal.

3. The method for online management of crew member assessment according to claim 2, characterized in that, The output ocean background noise specifically includes: Based on physical parameters, the theoretical attenuation coefficients of each frequency point at different propagation distances are calculated using the ocean acoustic attenuation equation, and the theoretical attenuation coefficients are arranged in order of frequency points to form a physical constraint weight matrix. The joint feature tensor is multiplied element-wise with the physical constraint weight matrix to obtain the physical weighted feature tensor. The physical weighted feature tensor is input into the physical constraint denoising network. During the training process, the loss function of the physical constraint denoising network includes a physical constraint term constructed based on the ocean acoustic attenuation equation. The physical constraint term is used to guide the network parameters to learn to extract noise components that match the physical parameters from the physical weighted feature tensor and output ocean background noise.

4. The method for online management of crew member assessment according to claim 1, characterized in that, S3 specifically includes: The pure crew member speech signal is subjected to spectrogram conversion to obtain spectrogram feature map, and syllable boundary detection is performed on the spectrogram feature map to segment it into several candidate term segments; Each candidate term fragment is input into a feature extraction network to extract the corresponding term acoustic feature vector. The cosine similarity between the acoustic feature vector of each term and the standard term prototype vector in the preset IMO standard term prototype library is calculated, and the term corresponding to the standard term prototype vector with the highest cosine similarity is taken as the recognition result; the term corresponding to the standard term prototype vector with the second highest cosine similarity is extracted as the easily confused candidate term. The recognition result is compared with the standard answer. If they are inconsistent, the similarity difference between the acoustic feature vector of the term and the standard term prototype vector corresponding to the recognition result, as well as the standard term prototype vector corresponding to the easily confused candidate term, is calculated. When the similarity difference is less than the preset threshold, the recognition result and the easily confused candidate term are marked as an easily confused term pair, and the term accuracy score and the confusion point information of the term containing the easily confused term pair are output.

5. The method for online management of crew member assessment according to claim 4, characterized in that, The step of inputting each candidate term fragment into a feature extraction network to extract the corresponding term acoustic feature vector specifically includes: The candidate term segments are segmented into multiple granularities to obtain both whole speech segments including complete terms and local speech segments including syllables. Mel frequency cepstral coefficients are extracted from the entire speech segment to obtain the overall feature vector, and linear predictive coding analysis is performed on the local speech segment to obtain the local feature vector reflecting the vocal tract resonance characteristics. Each frame in the local feature vector is arranged in chronological order to form a local temporal feature matrix; The overall feature vector and the local temporal feature matrix are concatenated and fused along the feature dimension, and the fused features are then subjected to an affine transformation to obtain the terminological acoustic feature vector.

6. The method for online management of crew member assessment according to claim 1, characterized in that, S4 specifically includes: Speech activity detection is performed on the clean crew voice signal to extract effective speech segments, and the average signal-to-noise ratio of the effective speech segments is calculated as the speech signal-to-noise ratio. The accuracy of a term is calculated by taking the proportion of correct terms to all terms in the statistical term accuracy score. The term accuracy rate is then obtained by looking up the corresponding term weight coefficient from a pre-defined weight mapping table based on the numerical range of the term accuracy rate. The sea state level is input into the sea state influence function, which outputs the corresponding sea state influence coefficient based on the pre-defined correspondence between the sea state level and the influence coefficient. The communication capability index is obtained by weighting and fusing the voice signal-to-noise ratio, term weight coefficient, and sea state influence coefficient.

7. The method for online management of crew member assessment according to claim 6, characterized in that, The output process of the sea state influence coefficient is as follows: Obtain the wave spectrum peak period and significant wave height corresponding to the sea state level, and calculate the wave energy density based on the wave spectrum peak period and the significant wave height; The wave energy density is input into the ship motion response function. The ship motion response function performs a weighted integral on the wave energy density according to the amplitude-frequency response operator corresponding to the ship type to obtain the ship heave motion amplitude. The amplitude of the ship's heave motion is input into the voice acquisition perturbation function. The voice acquisition perturbation function calculates the acceleration perturbation value generated at the microphone by the amplitude of the ship's heave motion based on the lever arm length between the microphone installation position and the ship's center of gravity. The acceleration disturbance value is input into the influence coefficient mapping network, which outputs the corresponding sea state influence coefficient based on the pre-calibrated correspondence between the acceleration disturbance value and the signal attenuation.

8. The method for online management of crew member assessment according to claim 1, characterized in that, S5 specifically includes: Based on the information on confusing terms, several knowledge point nodes associated with the confusing terms are extracted from the pre-set nautical terminology knowledge graph, and these knowledge point nodes are sorted according to their degree of association to obtain a knowledge point sequence. The communication capability index is compared with multiple preset capability level thresholds to determine the current capability level of the crew member, and the corresponding difficulty of the knowledge points to be strengthened is selected from the knowledge point sequence according to the capability level. The knowledge points to be strengthened are input into the training content matching library. Each training content in the training content matching library is pre-labeled with knowledge point tags and difficulty tags. By matching the knowledge points to be strengthened with the knowledge point tags, the training content that meets the difficulty tags is selected as candidate training content. The learning urgency coefficient is calculated based on the communication capability index, and the candidate training content is prioritized according to the learning urgency coefficient. The training content with the highest priority is selected to generate a push instruction, which is then sent to the crew terminal to present the corresponding online training course.

9. A method for online management of crew member assessment according to claim 8, characterized in that, The process involves matching the knowledge points to be reinforced with knowledge point tags to select training content that simultaneously meets the difficulty tags as candidate training content. Specifically, this includes: The knowledge points to be strengthened are input into the nautical terminology association network. Each node in the nautical terminology association network corresponds to a terminology knowledge point. The edges between nodes are pre-weighted according to the co-occurrence frequency of the terms in nautical operations. By traversing the nodes corresponding to the knowledge points to be strengthened, the adjacent nodes connected to the nodes and whose edge weights are greater than a preset threshold are extracted to form an extended knowledge point set. Each extended knowledge point in the extended knowledge point set is compared one by one with the knowledge point tags in the training content matching library. Training content whose knowledge point tags include extended knowledge points is selected to obtain the initial set of training content. Read the pre-marked difficulty tags of each training content in the initial selection training content set, compare the difficulty tags with the corresponding ability level thresholds of the knowledge points to be strengthened, and remove the training content whose difficulty tags are lower than the ability level thresholds to obtain candidate training content.

10. A seafarer assessment online management system, characterized in that, A method for implementing an online management system for crew member assessment as described in any one of claims 1-9 includes: The data acquisition module is used to collect the raw voice signals generated by the crew in the assessment scenario and simultaneously acquire the ship's navigation physical parameters at the corresponding time. The physical parameters include propeller speed, sea state level and ship speed. The noise reduction module inputs the original speech signal and physical parameters into the physical constraint noise reduction network. During the training phase, the physical constraint noise reduction network embeds the marine acoustic physics equation as a regularization term into the loss function and uses the physical parameters to guide the network to separate the marine background noise that conforms to the law of sound wave propagation from the original speech signal and output a clean crew speech signal. The terminology recognition module inputs the clean crew voice signal into the maritime terminology recognition network. The maritime terminology recognition network is based on prototype learning. By calculating the similarity between the voice features and the preset IMO standard term prototype vectors, it identifies the accuracy of terminology use in the crew's voice, marks easily confused term pairs, and outputs terminology accuracy scores and terminology confusion information. The comprehensive evaluation module calculates the crew's communication ability index based on the clean crew voice signal, terminology accuracy score and sea state level. The comprehensive evaluation algorithm integrates voice signal-to-noise ratio, terminology accuracy and sea state influence factor. The reporting and push module generates personalized performance evaluation reports based on communication ability index and terminology confusion information, and automatically matches and pushes corresponding online training course content based on the skill gaps in the report.

Images