Method and system for semantic understanding of water traffic based on machine learning and navigation assistance

Abstract

The application discloses a kind of water traffic semantic understanding and navigation method and system based on machine learning, it is related to water traffic communication field.The method obtains very high frequency channel audio and multi-base station time difference of arrival, text slice sequence is obtained by voice recognition;Based on the time difference, radio frequency reckoning coordinates are solved, and system registration coordinates are obtained by synchronously analyzing AIS data;Semantic intent and voice claim coordinates are extracted from text using algorithm;The spatial distance between radio frequency reckoning coordinates, system registration coordinates and voice claim coordinates is calculated respectively;The above distance is compared with the preset deviation threshold, spatial entity cross arbitration is executed, to determine the only credible spatial coordinates, and automatically navigation broadcast instruction is generated in combination with semantic intent.The subjective voice recognition result is corrected by objective physical data, effectively overcome the error recognition caused by crew's slip of the tongue or AI hallucination, greatly improve the safety and reliability of automatic navigation.

Description

Technical Field

[0001] This invention relates to the field of waterway communication technology, and in particular to a waterway semantic understanding and navigation aid method and system based on machine learning. Background Technology

[0002] Currently, maritime traffic management agencies, represented by VTS (Vessel Traffic Service), primarily communicate with vessels in their jurisdiction via VHF radiotelephone to implement traffic organization and distress safety communications. In recent years, the industry has attempted to introduce Automatic Speech Recognition (ASR) and Natural Language Processing (NLP) technologies to automatically extract semantic intent and entity elements (such as ship name and location coordinates) from ship communications in order to achieve automated risk assessment and navigational aid broadcasts.

[0003] However, existing semantic understanding solutions for water traffic have serious security risks. Their core flaw is that existing technologies simply use the "semantic location" extracted by ASR as factual evidence to input into the rule engine, relying entirely on subjective voice input and lacking a cross-verification mechanism with objective physical space.

[0004] In the complex and dangerous situations or severe sea conditions encountered in actual navigation, crew members are under high stress and are highly susceptible to verbally misreporting their own coordinates / reference buoys; or the ship's AIS (Automatic Identification System) equipment may malfunction or be tampered with, causing the track to drift. Current technology cannot autonomously determine whether "the crew misreported their voice or the physical sensors drifted." This disconnect between subjective semantics and objective physical location can lead to the system generating completely erroneous automatic navigation aids (such as broadcasting collision avoidance warnings to the wrong waters), thereby causing serious secondary navigation accidents. Summary of the Invention

[0005] This invention provides a machine learning-based method and system for semantic understanding and navigation aids in water traffic, in order to solve the problems existing in the background technology: the existing technology relies too much on subjective speech recognition results, and cannot autonomously identify and correct spatial entities when crew members misreport the location or AIS sensors fail, which leads to the failure of automatic navigation aid commands or even causes danger.

[0006] To achieve the aforementioned objectives, the technical solution adopted by this invention is: a machine learning-based method for semantic understanding and navigation aids in water traffic, executed by a computer device, comprising: S1. Obtain the time-series audio segment sequence in the VHF channel and the synchronized physical channel parameters including the time difference of arrival of multiple base stations. Based on the speech recognition model, perform speech transcription on the time-series audio segment sequence to obtain the text slice sequence. S2. Based on the radio positioning algorithm, the time difference of arrival of the multiple base stations is calculated to obtain the radio frequency estimated coordinates, and the real-time ship automatic identification system track data is parsed simultaneously to obtain the system registered coordinates; S3. Use a semantic extraction algorithm to extract the initial semantic intent and the voice claim coordinates containing location information from the text slice sequence; S4. Calculate the first spatial distance between the radio frequency estimated coordinates and the voice claim coordinates, and the second spatial distance between the system registered coordinates and the voice claim coordinates, respectively; S5. Compare the first spatial distance and the second spatial distance with a preset deviation threshold respectively, determine the unique credible spatial coordinates based on the comparison results, and generate an automatic navigation broadcast command by combining the unique credible spatial coordinates and the initial semantic intent.

[0007] Furthermore, after obtaining the text slice sequence, the process further includes: S11. Combine the base station identifier and radio frequency point in the physical channel parameters to generate a radio frequency fingerprint index, and use it as the key to write the text slice sequence into the semantic cache pool; S12. During the period before the lifecycle timer expires, extract the time-series text slices with the same radio frequency fingerprint index in the semantic cache pool to form a sequence of incomplete slot entities. S13. Input the incomplete slot entity sequence into the Long Short-Term Memory network model, and use the formula... Autoregressive decoding is performed to reconstruct and generate the initial semantic intent.

[0008] Furthermore, after acquiring the time-series audio segment sequence within the VHF channel, the method further includes: S14. Extract the signal-to-noise ratio data from the physical channel parameters and obtain the word-by-word confidence matrix output by the speech recognition model corresponding to the text slice sequence; S15. When the signal-to-noise ratio data is lower than the preset signal-to-noise ratio threshold, the attenuation formula is used. The weighted confidence level is obtained by performing a decay calculation on the base confidence level in the word-by-word confidence matrix. S16. When the weighted confidence level is lower than the preset intent threshold, the initial semantic intent in the text slice sequence is truncated and discarded.

[0009] Further, the step of calculating the radio frequency estimated coordinates by solving the time difference of arrival of the multiple base stations includes: S21. Multiply the preset electromagnetic wave propagation speed constant with the arrival time difference of the multiple base stations to obtain the fixed distance difference between the target ship and the main base station and the slave base station. S22. Using the physical coordinates of the master base station and the slave base station as geometric foci, construct multiple sets of hyperbolic positioning equations based on the fixed distance difference; S23. Solve the multiple sets of hyperbolic positioning equations using the Newton-Raphson iterative algorithm, and extract the coordinates of the effective geometric intersection points of the multiple sets of hyperbolic positioning equations in space as the radio frequency calculation coordinates.

[0010] Furthermore, determining the unique and reliable spatial coordinates based on the comparison results includes: S51, Using Formulas Calculate the tolerance radius and construct an RF error probability circle centered on the RF calculated coordinates; S52. When the second spatial distance is greater than the preset deviation threshold and the voice claim coordinates are located inside the radio frequency error probability circle, the voice claim coordinates are determined as the unique credible spatial coordinates. S53. When the second spatial distance is less than or equal to the preset deviation threshold and the voice claim coordinates are located outside the radio frequency error probability circle, the system registered coordinates are determined as the unique credible spatial coordinates.

[0011] Furthermore, the step of generating automatic navigation broadcast instructions by combining the unique trusted spatial coordinates and the initial semantic intent includes: S54. Based on the unique and reliable spatial coordinates, use the formula and formula The nearest encounter distance and collision time parameters of the target vessel are calculated; S55. Input the nearest encounter distance and the collision time parameters into a preset risk classification model to classify risk levels, and allocate the broadcast priority sequence of the automatic navigation broadcast command according to the risk level.

[0012] Furthermore, after allocating the broadcast priority sequence of the automatic navigation aid broadcast instructions according to the risk level, the method further includes: S56. When the current carrier energy value of the VHF channel is detected to be higher than a preset idle threshold, the priority flag bit in the automatic navigation broadcast command is parsed. S57. When the priority identifier corresponds to a normal service level, use the formula After calculating the backoff waiting time, the channel is re-detected; S58. When the priority flag corresponds to the emergency level, the underlying scheduling module is invoked to send a preemptive scheduling message to forcibly interrupt the ordinary communication link in the VHF channel and execute the broadcast.

[0013] Furthermore, after generating the automatic navigation broadcast command, the process also includes: S501. Collect audio slices of the target ship's response and convert them into text slices of the response. When the semantic similarity between the text slice of the response and the expected response slot is lower than a preset matching threshold, obtain manually annotated text. S502. Combine the text slice sequence with the manually annotated text to form incremental training sample pairs and load them into the background training engine; S503. Using the incremental training sample pair data, the backpropagation algorithm is executed based on the cross-entropy loss function to perform incremental parameter fine-tuning on the speech recognition model.

[0014] Another objective of this invention is to provide a machine learning-based semantic understanding and navigation aid system for water traffic. This system operates on a computer device with edge computing capabilities and includes: The data processing module is used to acquire the time-series audio segment sequence and physical channel parameters including the time difference of arrival of multiple base stations in the VHF channel, and to perform speech transcription on the time-series audio segment sequence based on the speech recognition model to obtain the text slice sequence. The coordinate calculation module is used to calculate the time difference of arrival of the multiple base stations based on the radio positioning algorithm to obtain the radio frequency estimated coordinates, and simultaneously parse the track data to obtain the system registered coordinates; The semantic extraction module is used to extract the initial semantic intent and speech claim coordinates from the text slice sequence; The distance comparison module is used to calculate the first spatial distance between the radio frequency estimated coordinates and the voice claim coordinates, and the second spatial distance between the system registered coordinates and the voice claim coordinates. The entity arbitration module is used to compare the first spatial distance and the second spatial distance with preset deviation thresholds, determine a unique credible spatial coordinate based on the comparison results, and generate an automatic navigation broadcast command by combining the unique credible spatial coordinate and the initial semantic intent.

[0015] The beneficial effects of this invention are as follows: 1. This invention introduces underlying radio physical parameters and uses purely physical radio frequency calculated coordinates as an absolute reference to perform multi-dimensional distance topology verification on the coordinates claimed by subjective speech. 2. In extreme cases where crew members make serious verbal slips due to nervousness, or when the ship's AIS sensors fail and the target deviation is extremely large, the system can intelligently arbitrate and correct the deviation based on the logic that the distance is small and does not exceed the threshold. This completely eliminates the risk of misbroadcasting of automatic navigation aids caused by the failure of a single information source, and greatly improves the safety and reliability of unmanned navigation aids on the water. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method flow provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the principle of spatial entity topology verification based on arrival time difference provided in an embodiment of the present invention; Figure 3 A flowchart of entity arbitration logic based on multi-source distance cross-calibration provided for embodiments of the present invention; Figure 4 This is a system hardware and functional module architecture diagram provided for an embodiment of the present invention. Detailed Implementation

[0017] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0018] like Figure 1 As shown, this application provides a machine learning-based method for semantic understanding and navigation aids in water traffic, executed by a computer device, including: Step S1: Obtain the time-series audio segment sequence in the VHF channel and the physical channel parameters containing the time difference of arrival of multiple base stations for synchronization matching. Based on the speech recognition model, the time-series audio segment sequence is transcribed into speech to obtain the text slice sequence. Step S2: Calculate the time difference of arrival of multiple base stations based on the radio positioning algorithm to obtain the radio frequency estimated coordinates, and simultaneously parse the real-time ship automatic identification system track data to obtain the system registered coordinates; Step S3: Use a semantic extraction algorithm to extract the initial semantic intent and the coordinates of the voice claim containing location information from the text slice sequence; Step S4: Calculate the first spatial distance between the radio frequency estimated coordinates and the voice claim coordinates, and the second spatial distance between the system registered coordinates and the voice claim coordinates, respectively; Step S5: Compare the first spatial distance and the second spatial distance with the preset deviation threshold respectively, determine the unique credible spatial coordinates based on the comparison results, and generate automatic navigation broadcast instructions by combining the unique credible spatial coordinates and the initial semantic intent.

[0019] Explanation: Step S1 serves to bridge the gap between analog speech signals and digital text. The voice emitted by the crew via the intercom (half-duplex mechanism) is discontinuous, i.e., a "time-sequence audio segment sequence." While capturing the audio, the system extracts the underlying objective physical parameters (the time difference between different base stations receiving the electromagnetic wave). Subsequently, a "text slice sequence" is output using a speech recognition model (such as Whisper).

[0020] Step S2 works by obtaining the ship's position using two different purely objective physical methods. First, it uses the time difference between the arrival times of electromagnetic waves at multiple base stations to inversely deduce the physical coordinates (RF-based coordinate calculation). Second, it reads the legally registered coordinates reported by the ship's onboard AIS sensors (system-registered coordinates). This provides an absolutely objective reference for subsequent anti-counterfeiting and error correction.

[0021] Step S3 aims to understand the subjective meaning of human language. The algorithm extracts the crew's subjective "initial semantic intent" (such as reporting location) and the "vocal claim coordinates" of the reference objects spoken by the crew from the transcribed text.

[0022] In steps S4 and S5, the system unifies subjective voice data and objective physical sensor data into the electronic chart coordinate system. It calculates the deviations of the subjective voice and AIS sensor data relative to the absolute physical reference (RF coordinates). When the subjective voice deviation is extremely large (greater than a preset deviation threshold) while the AIS deviation is extremely small, the system forcibly discards the voice recognition result and accepts the AIS data. Its technical effect lies in completely solving the fatal problem of crew members making mistakes or maliciously misreporting positions in harsh maritime environments, leading to AI generating incorrect navigational aids, thus achieving a highly robust error-proofing closed loop.

[0023] Example: The "Yun 123" vessel makes a call via a button amidst rough seas. The system receives audio with sea noise and records that the arrival time difference between base station 1 and base station 2 is 3 microseconds. The model transcribes this into text: "Traffic control, this is Yun 123, I am currently near buoy 12."

[0024] The system calculated the radio frequency coordinates (longitude 116.71E, latitude 23.32N) based on a 3-microsecond time difference; simultaneously, it read the system-registered coordinates reported by AIS (longitude 116.71E, latitude 23.33N). Both pointed to the waters of "Buoy No. 101".

[0025] The algorithm extracts the initial semantic intent as "ship position report" and extracts the voice claim coordinates as "buoy number 12" (the captain made a slip of the tongue due to nervousness, misreporting buoy number 101 as buoy number 12).

[0026] The system calculations show that the first spatial distance between the radio frequency estimated coordinates and the "buoy No. 12" subjectively reported by the captain is as high as 15 nautical miles; while the second spatial distance between the radio frequency estimated coordinates and the coordinates registered in the AIS system is only 0.1 nautical miles.

[0027] The preset deviation threshold is 1 nautical mile. Because the first spatial distance (15) > 1 and the second spatial distance (0.1) < 1, the system determines that the captain has made a verbal error. The system discards "buoy number 12", determines the AIS coordinates as the only reliable spatial coordinates (near buoy number 101), and generates a corrective instruction: "Mr. Yunyi 23, verify that your vessel is currently located at buoy number 101, not buoy number 12, please confirm".

[0028] Furthermore, after extracting the initial semantic intent and the speech claim coordinates containing location information, the process also includes: Step S11: Concatenate the base station identifier and radio frequency point in the physical channel parameters to generate a radio frequency fingerprint index, and write the text slice sequence into the semantic cache pool using the radio frequency fingerprint index as the key value; Step S12: During the period when the lifecycle timer associated with the semantic cache pool has not expired, extract the time-series text slices with the same radio frequency fingerprint index in the semantic cache pool to form the incomplete slot entity sequence. Step S13: Convert the incomplete slot entity sequence into a word vector sequence and input it into the Long Short-Term Memory network model to extract the hidden layer temporal features. ; Based on formula Autoregressive decoding is performed to generate a complete slot sequence, and the initial semantic intent is replaced with the reconstructed business intent generated based on the complete slot sequence, so as to generate automatic navigation broadcast instructions in the future.

[0029] Explanation: Steps S11 and S12 address the semantic fragmentation problem caused by half-duplex communication. The system uses "base station ID + transmission frequency" to generate a unique "radio frequency fingerprint index" as the primary key, and combines it with a time window (lifecycle timer) to aggregate fragmented texts intermittently transmitted by the same ship into a "sequence of incomplete slot entities".

[0030] Step S13 clarifies the network hierarchical connections of the artificial intelligence model. Long Short-Term Memory (LSTM) networks are adept at handling time-series sequences. In the formula, P represents the probability distribution. Predict the output for the current moment. For historical output, The input features are defined by , softmax is the normalized activation function, and W is the network weight matrix. 'b' represents the temporal features of the hidden layer, and 'b' is the bias term. The model calculates the probability distribution of each candidate word and repairs the incomplete intent. The technical effect is to achieve fragmented speech semantic self-healing and reconstruction across keystroke cycles.

[0031] Example: The captain sends a message three times via key presses: "Traffic control, this is..." / "...Transporter 123, requesting..." / "...Requesting tugboat." The system uses "Base Station_CH16" as the key and extracts these three segments into incomplete sequences within a 10-second lifecycle: "[Object: Traffic control] - [Subject: Transport] - [Action: Requesting tugboat]". After the LSTM model decodes, memorizes, and reassembles the formulas, the system reconstructs the complete business intent: [Subject: Transport 123, Target: Traffic control, Intent: Requesting tugboat rescue].

[0032] Furthermore, after acquiring the time-series audio segment sequence within the VHF channel, the process also includes: Step S14: Extract the signal-to-noise ratio data from the physical channel parameters and obtain the word-by-word confidence matrix output by the speech recognition model corresponding to the text slice sequence; Step S15: When the signal-to-noise ratio data is lower than the preset signal-to-noise ratio threshold, calculate the difference between the preset signal-to-noise ratio threshold and the signal-to-noise ratio data as an exponential decay variable; Using formula The calculated weighted values ​​are used as the weighted confidence scores, where This represents the base confidence level in the word-by-word confidence matrix. For the preset attenuation coefficient, The preset signal-to-noise ratio threshold is used, and SNR is the signal-to-noise ratio data. Step S16: When the weighted confidence level is lower than the preset intent threshold, the initial semantic intent is truncated and discarded.

[0033] Explanation: Steps S14 to S16 introduce a confidence reduction penalty mechanism based on the underlying physical signal. The principle is to avoid blindly trusting the output of the speech recognition model and instead forcibly introduce physical noise floor (SNR) as a penalty factor. The lower the SNR, the smaller the negative exponent of e, resulting in an exponential decrease in the calculated "weighted confidence." This technical effect eliminates the risk of "text illusion" and subsequent erroneous broadcasts caused by the speech recognition model under poor channel conditions, fundamentally addressing the issue from the physical signal level.

[0034] Example: The "Yun 123" vessel transmitted a message accompanied by significant wind noise. The signal-to-noise ratio (SNR) was only 5 dB (below the threshold). =15dB). The speech recognition model forcibly output the hallucinatory text "entering the wind farm" amidst background noise, and the model's own base confidence level was 15dB. The confidence level is as high as 0.9. Substituting the system into the decay formula, due to the extremely large exponential decay variable (15-5=10), the calculated weighted confidence level is... The value dropped sharply to 0.2. Since 0.2 < 0.6 (the preset intent threshold), the system determined that the text was an AI illusion and directly truncated and discarded it, thus avoiding the issuance of a false alarm.

[0035] Furthermore, radio frequency (RF) coordinates are obtained by calculating the time difference of arrival (TDOA) of multiple base stations based on radio positioning algorithms, including: Step S21: Calculate the fixed distance difference between the target ship and the main base station and the slave base station by multiplying the preset electromagnetic wave propagation speed constant with the arrival time difference of multiple base stations. Step S22: Using the actual physical coordinates of the master base station and the slave base station as the focus, construct multiple sets of hyperbolic positioning equations based on the fixed distance difference; Step S23: Solve multiple sets of hyperbolic positioning equations using the Newton-Raphson iterative algorithm, and extract the coordinates of the effective geometric intersection points as the radio frequency calculation coordinates.

[0036] Explanation: Steps S21 to S23 belong to the classic TDOA (Time Difference of Arrival) positioning mathematical model. The speed of electromagnetic waves (speed of light) multiplied by the time difference can be converted into a distance difference. According to the geometric definition, the trajectory of points whose distances to two focal points differ by a constant is a hyperbola. By solving multiple sets of nonlinear hyperbolic equations using Newton's iterative algorithm, the source of the transmission can be accurately located.

[0037] Example: The speed of electromagnetic waves is approximately 300 meters per microsecond. The time difference between the reception of voice signals at one base station and two base stations is 3 microseconds. Multiplying this by the base station yields a fixed distance difference of 900 meters between the ship and the two base stations. The system constructs multiple sets of hyperbolic equations with the two base stations as the focal points. Using a Newton-Raphson iterative algorithm, it quickly approximates these equations to determine the effective geometric intersection coordinates of multiple hyperbolas on the sea surface as (116.71E, 23.32N), which are the precise radio frequency calculated coordinates.

[0038] Furthermore, based on the comparison results, the voice claim coordinates or system registered coordinates that are close in distance and do not exceed a preset deviation threshold are determined as unique reliable spatial coordinates, including: Step S51: Using the formula Calculate the tolerance radius and construct an RF error probability circle centered on the RF calculated coordinates, where c is the speed of light. The preset time synchronization error is P, where P is the probability confidence level. Step S52: When the second spatial distance is greater than the preset deviation threshold and the voice claim coordinates are located inside the radio frequency error probability circle, discard the system-registered coordinates and extract the voice claim coordinates as the only reliable spatial coordinates. Step S53: When the second spatial distance is less than or equal to the preset deviation threshold and the voice claim coordinates are located outside the radio frequency error probability circle, discard the voice claim coordinates and extract the system registered coordinates as the only reliable spatial coordinates.

[0039] Explanation: Step S51 considers the multipath effect of radio waves at sea and uses the Gaussian error distribution formula to define the maximum physical measurement error boundary (RF error probability circle) that the system can tolerate. Step S52 describes the arbitration scenario of "AIS equipment malfunction and tampering": if the AIS deviates greatly, but the captain's spoken coordinates fall perfectly within the physical RF circle, the system accepts the captain's voice. Step S53 describes the arbitration scenario of "the captain makes a serious slip of the tongue": if the AIS closely matches the physical RF circle, and only the captain's spoken coordinates deviate greatly, the system accepts the AIS data.

[0040] Example of integration: Substituting the speed of light c with the synchronization error (0.1 microseconds) The tolerance radius is calculated to be 50 meters. The system draws a 50-meter radio frequency error probability circle around the radio frequency coordinates. If the AIS equipment of "Yun 123" malfunctions, causing the coordinates to drift to more than 10 nautical miles (second spatial distance > deviation threshold), but the captain reports "at buoy number 101" and falls exactly within this 50-meter circle, step S52 is triggered, and the system accepts the voice coordinates. Conversely, if step S53 is triggered, the AIS coordinates are accepted as in step S5.

[0041] Furthermore, by combining unique, reliable spatial coordinates and the initial semantic intent, automatic navigation broadcast instructions are generated, including: Step S54: Obtain the relative motion vector parameters between the target ship and surrounding ships based on the unique and reliable spatial coordinates; Using formula Calculate the relative velocity modulus and use the formula Calculate the nearest meeting distance ; In the formula, and These represent the absolute speed components of the target ship along the X and Y axes, respectively, calculated by the system in a preset two-dimensional relative motion coordinate system. and These represent the extracted absolute speed components of surrounding vessels along the X and Y axes in the same coordinate system, respectively. formula In the vector, R is the initial relative position vector. This is the relative velocity modulus calculated previously; When the nearest encounter distance is less than the preset safety distance, use the formula The collision time parameters were calculated. ; Step S55: Extract the nearest encounter distance and collision time parameters to classify the risk level, and allocate the broadcast priority sequence of automatic navigation broadcast instructions according to the risk level.

[0042] Explanation: Steps S54 to S55 are classic geometric mathematical models for maritime collision avoidance. Relative velocity is obtained by subtracting vectors and taking their modulus. DCPA is the nearest encounter distance, and TCPA is the time to reach the most dangerous encounter point. The system queues and prioritizes instructions based on both temporal and spatial urgency.

[0043] Example: The system uses the unique, reliable spatial coordinates of "Yun 123" as the origin and detects a fishing vessel crossing the channel. Substituting the speed and heading vectors of both vessels, the DCPA (Distance Distance Aspect) is calculated to be only 0.1 nautical miles (extremely dangerous). Because 0.1 nautical miles < 0.5 nautical miles (the preset safe distance), the TCPA (Collision Time Aspect) formula is triggered, calculating a collision time parameter of 3 minutes. Based on this, the system classifies it as the highest risk and assigns it an absolute priority broadcast sequence.

[0044] Furthermore, after allocating the broadcast priority sequence of automatic navigation aid broadcast instructions according to risk level, it also includes: Step S56: Detect the current carrier energy value of the VHF channel. When the current carrier energy value is higher than the preset channel idle threshold, parse the priority flag bit in the automatic navigation aid broadcast command. Step S57: When the priority flag corresponds to the ordinary service level, use the formula... After calculating the backoff waiting time, the collision is re-detected, where n is the number of collision backoff attempts. Based on the time slot length; Step S58: When the priority flag corresponds to the emergency level, the underlying scheduling module is invoked to send a preemptive scheduling message to forcibly interrupt the ordinary communication link in the VHF channel and execute the broadcast.

[0045] Explanation: Steps S56 to S58 introduce collision control and carrier sensing mechanisms. For ordinary broadcasts (such as weather forecasts), a binary exponential backoff algorithm is used (calculated...). Random waiting is used to avoid signal interference; while the highest-level instructions involving human safety directly issue underlying signaling to forcibly cut off other communications and monopolize the channel for broadcasting. The technical effect is to ensure the absolute unobstructed flow of life-saving channels.

[0046] Example scenario: When the system is preparing to broadcast a collision avoidance warning to "Yun 123", it detects that channel CH16 is being occupied by two other tugboats chatting (high carrier energy value). Because the priority flag of the collision avoidance warning is "emergency level of distress", the system does not execute the backoff formula, but instead directly sends a preemption scheduling message, forcibly cutting off the tugboat's communication and broadcasting at high power: "Yun 123, extremely dangerous, turn around immediately!"

[0047] Furthermore, after generating the automatic navigation broadcast command, it also includes: Step S501: Collect the actual audio slices of the target ship's response, call the speech recognition model to convert them into response text slices, and calculate the semantic similarity index between them and the expected response slot; Step S502: When the semantic similarity index is lower than the preset matching threshold and a manual intervention correction signal is received, extract the corresponding manually annotated text that has passed manual verification. Step S503: Combine the text slice sequence with manually annotated text to form incremental training sample pairs, and perform incremental parameter fine-tuning on the speech recognition model based on the cross-entropy loss function.

[0048] Explanation: Steps S501 to S503 rigorously clarify the closed-loop process of model training. When AI recognition fails, causing the process to stall, the paired data generated by manual intervention and correction will be fed into the deep learning framework. Gradients will be calculated using cross-entropy loss and backpropagation will be used to update the weights. The technical effect is that the system becomes more accurate with use, gradually understanding the dialects and slang of specific waters.

[0049] Example: After the system broadcasts a warning, the captain of the ship "Yun 123" replies "Zhi la" in Chaoshan dialect. The system fails to recognize this as the expected slot "Should be received," resulting in an extremely low semantic similarity index. The on-duty personnel intervene and manually input the annotated text: "meaning received." The system then feeds the model with incremental training samples composed of the dialect audio and the correct text. After fine-tuning using cross-entropy loss, the model completes its self-evolution.

[0050] The technical solutions in the above embodiments will be further described below with reference to the accompanying drawings.

[0051] like Figure 1 The diagram shown is a schematic representation of the method flow provided in an embodiment of the present invention. The present invention provides a method for semantic understanding and navigation aids in maritime traffic based on machine learning. This method is automatically executed by computer equipment (such as an edge computing server deployed within a maritime private network) and specifically includes the following steps: Step S1: Obtain the time-series audio segment sequence in the VHF channel and the physical channel parameters containing the time difference of arrival of multiple base stations for synchronous matching. Based on the speech recognition model, perform speech transcription on the time-series audio segment sequence to obtain the text slice sequence.

[0052] Step S2: Calculate the time difference of arrival of multiple base stations based on the radio positioning algorithm to obtain the radio frequency estimated coordinates, and simultaneously parse the real-time ship automatic identification system track data to obtain the system registered coordinates.

[0053] Step S3: Use a semantic extraction algorithm to extract the initial semantic intent and the voice claim coordinates containing location information from the text slice sequence.

[0054] Step S4: Calculate the first spatial distance between the radio frequency estimated coordinates and the voice claimed coordinates, and the second spatial distance between the system registered coordinates and the voice claimed coordinates.

[0055] Step S5: Compare the first spatial distance and the second spatial distance with the preset deviation threshold respectively. Based on the comparison results, determine the radio frequency calculated coordinates or system registered coordinates with smaller distances that do not exceed the preset deviation threshold as the unique reliable spatial coordinates. Combine the unique reliable spatial coordinates and the initial semantic intent to generate automatic navigation broadcast instructions.

[0056] To further illustrate the technical solution and actual technical effects of the present invention, the following is combined with... Figure 2 and Figure 3 The diagram shown illustrates the logic principle of cross-arbitration of three-dimensional spatial entity topology. Taking a real navigation scenario of the "Yun 123" vessel in a certain waterway as an example, it details the automated navigation aid closed-loop execution process of this system when facing extreme situations. (1) Scene triggering and data acquisition (corresponding steps S1-S3) Under severe weather conditions, the captain of the "Yun 123" vessel pressed the intercom on the VHF CH16 channel to call the traffic control center.

[0057] Physical layer data acquisition: The converged communication gateway acquires the audio signal accompanied by sea wave noise and simultaneously records the arrival time difference of the radio frequency signal received by the Nan'ao main base station and the Haojiang secondary base station as 3 microseconds.

[0058] Objective coordinate calculation: The computer equipment used the aforementioned 3-microsecond time difference to calculate the absolute physical location, i.e., the radio frequency calculated coordinates, using a hyperbolic positioning algorithm, as (longitude 116.71E, latitude 23.32N); simultaneously, it read the system-registered coordinates reported by the ship's AIS sensor via the interface, which were (longitude 116.71E, latitude 23.33N). Both coordinates pointed to the area of ​​"Bus No. 101" on the electronic nautical chart.

[0059] Subjective semantic extraction: The speech recognition model transcribes the audio into text: "Traffic control, this is 10123, I am currently near buoy 12." The semantic extraction algorithm extracts the initial semantic intent as "[ship position report]" and extracts the subjective voice claim coordinates as "[buoy 12]" (here the captain made a slip of the tongue due to high tension, misreporting buoy 101 as buoy 12).

[0060] (2) Spatial cross-arbitration and correction broadcast (corresponding steps S4-S5) Quantitative Spatial Deviation: The computer equipment maps the three sets of coordinates mentioned above onto the electronic chart plane. Calculations show that the first spatial distance between the objective radio frequency (RF) estimated coordinates and the subjectively reported "Buoy No. 12" is as high as 15 nautical miles; the second spatial distance between the objective system-registered coordinates (AIS) and the subjectively reported "Buoy No. 12" is also as high as 15 nautical miles. However, the distance error between the RF estimated coordinates and the system-registered coordinates is extremely small (only about 0.1 nautical miles).

[0061] Entity arbitration logic: The system's preset deviation threshold is 1 nautical mile. Since both the first spatial distance (15 nautical miles) and the second spatial distance are much greater than the preset deviation threshold of 1 nautical mile, the system triggers the spatial entity cross-arbitration mechanism. System determination: Since the underlying physical radio frequency positioning and AIS sensor positioning are highly consistent, it must be that the captain made a serious subjective error.

[0062] Error-prevention closed-loop broadcast: The system forcibly discards the subjectively extracted "buoy number 12" and determines the objective AIS system registered coordinates as the only reliable spatial coordinates (i.e., near buoy number 101). In conjunction with the intent of the [Vessel Position Report], the system automatically intercepts instructions that may be sent to the wrong waters and generates a corrective instruction, which is then broadcast to the base station: "Vessel 101, the system has verified that your vessel is currently located at buoy number 101, not buoy number 12. Please confirm."

[0063] As can be seen from the above specific application examples, the present invention completely eliminates the risk of mis-broadcasting of automatic navigation aids caused by a single human subjective error (mispronounced location) or a single sensor failure (AIS drift), and achieves a highly robust error prevention closed loop.

[0064] like Figure 4 The diagram shown illustrates the system hardware and functional module architecture provided in an embodiment of the present invention. This embodiment also provides a machine learning-based maritime traffic semantic understanding and navigation aid system. This system runs on a computer device with edge computing capabilities (such as a maritime safety administration intranet business front-end machine).

[0065] The computer device includes, at the hardware level, a processor, memory, and a communication interface 100. The memory stores computer instructions that can be executed by the processor. At the logical functional level, the system specifically includes the following functional modules: Data processing module 200: Coupled with communication interface 100, it is used to acquire the time-series audio segment sequence in the VHF channel and the physical channel parameters containing the time difference of arrival of multiple base stations for synchronous matching, and to perform speech transcription on the time-series audio segment sequence based on the speech recognition model to obtain the text slice sequence. Coordinate calculation module 300: Called by the processor, it is used to calculate the time difference of arrival of multiple base stations based on the radio positioning algorithm to obtain the radio frequency estimated coordinates, and simultaneously parse the real-time ship automatic identification system track data to obtain the system registered coordinates; Semantic extraction module 400: used to extract the initial semantic intent and the coordinates of the speech claim containing location information from the text slice sequence using a semantic extraction algorithm; Distance comparison module 500: responsible for quantifying and calculating the first spatial distance between the radio frequency estimated coordinates and the voice claim coordinates, and the second spatial distance between the system registered coordinates and the voice claim coordinates in the constructed two-dimensional coordinate system; The entity arbitration module 600 is coupled to the communication interface 100 to perform low-level control. It is used to compare the first spatial distance and the second spatial distance with the preset deviation threshold respectively. Based on the comparison result, the radio frequency calculated coordinates or system registered coordinates with smaller distances and not exceeding the preset deviation threshold are determined as unique reliable spatial coordinates. The module combines the unique reliable spatial coordinates and the initial semantic intent to generate automatic navigation aid broadcasting instructions for forced correction.

[0066] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A machine learning-based method for semantic understanding and navigation aids in water traffic, characterized in that, Performed by computer devices, including: S1. Obtain the time-series audio segment sequence in the VHF channel and the synchronized physical channel parameters including the time difference of arrival of multiple base stations. Based on the speech recognition model, perform speech transcription on the time-series audio segment sequence to obtain the text slice sequence. S2. Based on the radio positioning algorithm, the time difference of arrival of the multiple base stations is calculated to obtain the radio frequency estimated coordinates, and the real-time ship automatic identification system track data is parsed simultaneously to obtain the system registered coordinates; S3. Use a semantic extraction algorithm to extract the initial semantic intent and the voice claim coordinates containing location information from the text slice sequence; S4. Calculate the first spatial distance between the radio frequency estimated coordinates and the voice claim coordinates, and the second spatial distance between the system registered coordinates and the voice claim coordinates, respectively; S5. Compare the first spatial distance and the second spatial distance with a preset deviation threshold respectively, determine the unique credible spatial coordinates based on the comparison results, and generate an automatic navigation broadcast command by combining the unique credible spatial coordinates and the initial semantic intent.

2. The water traffic semantic understanding and navigation aid method based on machine learning according to claim 1, characterized in that, After obtaining the text slice sequence, the process further includes: S11. Combine the base station identifier and radio frequency point in the physical channel parameters to generate a radio frequency fingerprint index, and use it as the key to write the text slice sequence into the semantic cache pool; S12. During the period before the lifecycle timer expires, extract the time-series text slices with the same radio frequency fingerprint index in the semantic cache pool to form a sequence of incomplete slot entities. S13. Input the incomplete slot entity sequence into the Long Short-Term Memory network model, and use the formula... Autoregressive decoding is performed to reconstruct and generate the initial semantic intent.

3. The machine learning-based semantic understanding and navigation aid method for water traffic according to claim 1, characterized in that, After acquiring the time-series audio segment sequence within the VHF channel, the method further includes: S14. Extract the signal-to-noise ratio data from the physical channel parameters and obtain the word-by-word confidence matrix output by the speech recognition model corresponding to the text slice sequence; S15. When the signal-to-noise ratio data is lower than the preset signal-to-noise ratio threshold, the attenuation formula is used. The weighted confidence level is obtained by performing a decay calculation on the base confidence level in the word-by-word confidence matrix. S16. When the weighted confidence level is lower than the preset intent threshold, the initial semantic intent in the text slice sequence is truncated and discarded.

4. The machine learning-based semantic understanding and navigation aid method for water traffic according to claim 1, characterized in that, The step of calculating the radio frequency (RF) coordinates by solving the time difference of arrival of the multiple base stations includes: S21. Multiply the preset electromagnetic wave propagation speed constant with the arrival time difference of the multiple base stations to obtain the fixed distance difference between the target ship and the main base station and the slave base station. S22. Using the physical coordinates of the master base station and the slave base station as geometric foci, construct multiple sets of hyperbolic positioning equations based on the fixed distance difference; S23. Solve the multiple sets of hyperbolic positioning equations using the Newton-Raphson iterative algorithm, and extract the coordinates of the effective geometric intersection points of the multiple sets of hyperbolic positioning equations in space as the radio frequency calculation coordinates.

5. The machine learning-based semantic understanding and navigation aid method for water traffic according to claim 1, characterized in that, The determination of unique and reliable spatial coordinates based on the comparison results includes: S51, Using Formulas Calculate the tolerance radius and construct an RF error probability circle centered on the RF calculated coordinates; S52. When the second spatial distance is greater than the preset deviation threshold and the voice claim coordinates are located inside the radio frequency error probability circle, the voice claim coordinates are determined as the unique credible spatial coordinates. S53. When the second spatial distance is less than or equal to the preset deviation threshold and the voice claim coordinates are located outside the radio frequency error probability circle, the system registered coordinates are determined as the unique credible spatial coordinates.

6. The water traffic semantic understanding and navigation aid method based on machine learning according to claim 1, characterized in that, The step of generating automatic navigation broadcast instructions by combining the unique trusted spatial coordinates and the initial semantic intent includes: S54. Based on the unique and reliable spatial coordinates, use the formula and formula The nearest encounter distance and collision time parameters of the target vessel are calculated; S55. Input the nearest encounter distance and the collision time parameters into a preset risk classification model to classify risk levels, and allocate the broadcast priority sequence of the automatic navigation broadcast command according to the risk level.

7. The machine learning-based semantic understanding and navigation aid method for water traffic according to claim 6, characterized in that, After allocating the broadcast priority sequence of the automatic navigation aid broadcast instructions according to the risk level, the method further includes: S56. When the current carrier energy value of the VHF channel is detected to be higher than a preset idle threshold, the priority flag bit in the automatic navigation broadcast command is parsed. S57. When the priority identifier corresponds to a normal service level, use the formula After calculating the backoff waiting time, the channel is re-detected; S58. When the priority flag corresponds to the emergency level, the underlying scheduling module is invoked to send a preemptive scheduling message to forcibly interrupt the ordinary communication link in the VHF channel and execute the broadcast.

8. The machine learning-based semantic understanding and navigation aid method for water traffic according to claim 1, characterized in that, After generating the automatic navigation broadcast command, the process also includes: S501. Collect audio slices of the target ship's response and convert them into text slices of the response. When the semantic similarity between the text slice of the response and the expected response slot is lower than a preset matching threshold, obtain manually annotated text. S502. Combine the text slice sequence with the manually annotated text to form incremental training sample pairs and load them into the background training engine; S503. Using the incremental training sample pair data, the backpropagation algorithm is executed based on the cross-entropy loss function to perform incremental parameter fine-tuning on the speech recognition model.

9. A system applied to the machine learning-based semantic understanding and navigation aid method for water traffic as described in any one of claims 1 to 8, characterized in that, The system runs on a computer device with edge computing capabilities and includes: The data processing module is used to acquire the time-series audio segment sequence and physical channel parameters including the time difference of arrival of multiple base stations in the VHF channel, and to perform speech transcription on the time-series audio segment sequence based on the speech recognition model to obtain the text slice sequence. The coordinate calculation module is used to calculate the time difference of arrival of the multiple base stations based on the radio positioning algorithm to obtain the radio frequency estimated coordinates, and simultaneously parse the track data to obtain the system registered coordinates; The semantic extraction module is used to extract the initial semantic intent and speech claim coordinates from the text slice sequence; The distance comparison module is used to calculate the first spatial distance between the radio frequency estimated coordinates and the voice claim coordinates, and the second spatial distance between the system registered coordinates and the voice claim coordinates. The entity arbitration module is used to compare the first spatial distance and the second spatial distance with preset deviation thresholds, determine a unique credible spatial coordinate based on the comparison results, and generate an automatic navigation broadcast command by combining the unique credible spatial coordinate and the initial semantic intent.

Images