Intelligent seat control system based on AI voice and star flash technology

Abstract

This invention provides an intelligent seat control system based on AI voice and StarFlash technology, relating to the field of artificial intelligence technology. The system includes: directional acquisition of target occupant voice signals; noise reduction processing using an improved adaptive noise reduction algorithm to obtain voice segments with noise feature labels; AI adaptive processing to obtain ordinary parsed seat control transactions; fuzzy intent disambiguation processing of the ordinary parsed seat control transactions to obtain disambiguated parsed seat control transactions; classification processing to generate a control transaction instruction sequence; real-time acquisition of feedback data during the execution of the control transaction instruction sequence; construction of multi-scenario safety interlock rules and multi-level anomaly degradation strategies; primary safety interlock control of the feedback data based on the multi-scenario safety interlock rules and real-time driving status data; and secondary safety interlock control of the seat based on the multi-level anomaly degradation strategies, thereby improving the safety of seat adjustment.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically an intelligent seat control system based on AI voice and Star Flash technology. Background Technology

[0002] With the upgrading of automotive intelligence, in-vehicle seat control is developing towards convenience and safety, but existing systems have many pain points. In terms of control, physical operation is distracting, basic voice control is affected by in-vehicle noise, has low recognition accuracy and cannot resolve ambiguous commands, making it difficult to understand the natural expressions and true intentions of passengers, and unable to make intelligent adaptive adjustments based on passenger preferences and changes in the scene. Therefore, there is a need for an intelligent seat control solution that can improve the accuracy of voice interaction recognition and understand passenger intentions while maintaining the safety constraints of in-vehicle control, thereby reducing repeated debugging and repetitive operations.

[0003] In vehicle communication, traditional technologies such as Bluetooth and Wi-Fi suffer from high latency and weak interference resistance, easily leading to abnormal command transmission. In terms of security, their protection mechanisms are simplistic, lacking multi-scenario security interlocks and comprehensive anomaly degradation strategies, posing security risks. StarFlash technology boasts low latency and high reliability, while AI adaptive voice recognition can overcome noise limitations; combining the two can solve existing pain points. Therefore, there is an urgent need to develop an intelligent seat control system that integrates both technologies, enabling convenient, precise, and safe seat adjustments to meet the demands of intelligent vehicle systems. Summary of the Invention

[0004] In order to solve the technical problem of reducing repeated debugging and repetitive operation by improving the accuracy of voice interaction recognition and understanding passenger intentions while maintaining vehicle control safety constraints, the present invention aims to provide an intelligent seat control system based on AI voice and Star Flash technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent seat control system based on AI voice and Star Flash technology, the system comprising: a voice acquisition and processing module, an AI adaptive voice recognition module, a Star Flash wireless communication module, a seat control execution module, and a safety interlock module; The voice acquisition and processing module is used to acquire the voice signal of the target occupant in a directional manner, and to perform noise reduction processing on the voice signal of the target occupant according to the improved adaptive noise reduction algorithm to obtain voice segments with noise feature labels. The AI ​​adaptive speech recognition module is used to perform AI adaptive processing on the speech segments with noise feature labels to obtain ordinary parsed seat control transactions; and to perform fuzzy intent disambiguation processing on the ordinary parsed seat control transactions to obtain the corresponding disambiguated parsed seat control transactions. The StarScan wireless communication module is used to classify and process the disambiguation parsing seat control transactions and generate corresponding control transaction instruction sequences. The seat control execution module is used to execute the control transaction instruction sequence and collect feedback data in real time during the execution of the control transaction instruction sequence; The safety interlock module is used to construct multi-scenario safety interlock rules and multi-level anomaly degradation strategies; based on the multi-scenario safety interlock rules and real-time driving status data, it performs first-level safety interlock control on the feedback data; based on the multi-level anomaly degradation strategies, it performs second-level safety interlock control on the seat; based on the two-level safety interlock control, the safety of seat adjustment is improved.

[0006] Preferably, the process of obtaining speech segments with noise feature labels includes: A microphone array is used to spatially directionally acquire occupant speech. Beamforming algorithm is used to lock the direction of the target sound source and suppress environmental noise in non-target areas. Engine noise echo, wind noise echo, and audio system echo are used as reference signals. Adaptive echo cancellation is performed using a normalized least mean square algorithm, and the echo suppression weights are iteratively updated. Based on segmented spectral subtraction and noise estimation algorithms, the target occupant speech signal after adaptive echo cancellation is denoised to obtain the corresponding speech segments. The corresponding signal-to-noise ratio, noise power spectrum, and noise interference type are calculated in real time. The corresponding signal-to-noise ratio, noise power spectrum, and noise interference type are encoded as noise feature tags and bound to the denoised speech segments for output, forming speech segments with noise feature tags.

[0007] Preferably, the process of obtaining a normal parsed seat control transaction includes: A three-layer model architecture is adopted, including an acoustic model, a language model, and a BIO slot labeling model. Speech segments with noise feature labels are input into the acoustic model, and the recognition threshold and feature weights are dynamically adjusted according to the noise feature labels to convert noise-robust speech segments into text data. The text data is input into the language model, and the text is structured and parsed using statistical rules and lightweight semantic coding. Then, the corresponding adjustment actions, adjustment objects, adjustment amplitudes, and target modes are extracted according to the BIO slot labeling model, and legality verification and format standardization are performed to obtain ordinary parsed seat control transactions.

[0008] Preferably, the process of obtaining the corresponding disambiguation resolution seat control transaction includes: Weakly supervised samples are constructed based on the target occupant's historical adjustment records, habit range, and preference angle data, thereby building a seat control preference model for the target occupant. Based on this seat control preference model, fuzzy instructions for ordinary parsing seat control transactions are disambiguated. A confidence score for the corresponding disambiguation result is calculated based on a confidence assessment model. When the confidence score is greater than or equal to a confidence threshold, the corresponding disambiguated parsing seat control transaction is directly obtained. When the confidence score is less than the confidence threshold, a minimum clarification interaction is triggered, thereby obtaining the corresponding disambiguated parsing seat control transaction.

[0009] Preferably, the process of generating the corresponding control transaction instruction sequence includes: A low-latency, short-range wireless link within the vehicle is established based on the NearLink protocol, and corresponding node pairing and link synchronization are configured. Seat control transactions are categorized into high-priority and normal-priority commands based on disambiguation analysis of their security level and urgency. QoS is configured for high-priority commands, and a resource unit dynamic scheduling algorithm is used to prioritize the allocation of time-frequency resources and transmission time slots. A sequence number, checksum, and retransmission flag are added to each command, and an acknowledgment-retransmission mechanism is constructed to correct abnormal commands. The commands are then sorted according to priority level and time order to form a corresponding control transaction command sequence.

[0010] Preferably, the process of constructing multi-scenario security interlocking rules includes: Based on the vehicle safety knowledge graph, the correlation strength between seat safety status and various driving scenarios is calculated. By traversing the network and accumulating the contribution scores of multi-source seat safety status parameters to different driving scenarios, several candidate driving scenarios are selected. Historical driving scenarios and seat control records are retrieved, and historical seat safety status features of similar driving scenarios are extracted. The similarity between the seat safety status parameters of the current candidate driving scenario and the historical seat safety status features of the corresponding driving scenario is calculated. All candidate driving scenarios are traversed according to their contribution scores. If the parameter similarity is less than the confirmation threshold, the scenario is not considered a candidate driving scenario. If the parameter similarity is greater than or equal to the confirmation threshold, the safety interlock constraint type of the corresponding candidate driving scenario is determined. The seat control entity associated with the corresponding safety interlock constraint type is combined with the specific adjustment mechanism and safety interlock adjustment action type corresponding to the target seat status parameters to clarify the executable safety restriction rules, thereby forming multi-scenario safety interlock rules.

[0011] Preferably, the process of constructing a multi-level anomaly degradation strategy includes: The system collects the operating status and fault codes between modules in real time, calculates the correlation weight between abnormal indicators of each module and the system fault level, sets abnormal judgment criteria through historical abnormal cases, and generates corresponding abnormal degradation strategies. Based on the correlation weight, the system selects the corresponding abnormal degradation strategies. For each abnormal degradation strategy, the system formulates control schemes to maintain the current position, roll back to a safe posture, or switch local control, and binds corresponding trigger conditions and execution processes to form corresponding multi-level abnormal degradation strategies.

[0012] Preferably, the process of performing first-level safety interlock control on the feedback data based on the multi-scenario safety interlock rules and real-time driving status data includes: The system calculates the correlation strength between the current real-time driving status data and the corresponding feedback data for the primary safety interlock control. Based on the correlation strength, it selects safety interlock rules for the current candidate driving scenario. According to the seat adjustment object and actuator associated with the safety constraint rules, it matches the adjustment limitation instructions in the corresponding feedback data. The adjustment limitation instructions include seat motor voltage limitation instructions, seat motor current limitation instructions, adjustment direction limitation instructions, and adjustment stroke parameter limitation instructions, thereby realizing primary safety interlock control and outputting the constraint status.

[0013] Preferably, the process of performing secondary safety interlock control on the seat according to the multi-level anomaly degradation strategy includes: The system monitors the operating status of each module during the execution of control transaction instruction sequences, calculates the correlation strength between the corresponding module's abnormal indicators and the secondary safety interlock control of each level of abnormal degradation strategy, sets correlation weights based on the correlation strength of the secondary safety interlock control, and selects the abnormal degradation strategy with the highest adaptability. Based on the control mode and safety posture associated with the abnormal degradation strategy, the system executes the corresponding control scheme, thereby realizing the secondary safety interlock control of the current seat operating status.

[0014] The present invention also provides a computer-readable storage medium storing a computer program, the computer program being executed by a processor of the intelligent seat control system based on AI voice and star-flash technology.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. Solving the problem of complex noise interference in the vehicle, making voice acquisition more accurate: By using microphone array directional acquisition and beamforming sound source locking technology in the voice acquisition and processing module, combined with an improved adaptive noise reduction algorithm, non-target noise such as engine noise, wind noise, and audio echo can be effectively suppressed. At the same time, adaptive echo cancellation is achieved through the normalized least mean square algorithm, and deep noise reduction is completed by combining segmented spectral subtraction and noise estimation algorithms. Finally, the voice segments with noise feature labels are output, providing high-quality input for subsequent recognition, avoiding recognition misjudgment caused by noise, and greatly improving the anti-interference capability of voice acquisition.

[0016] 2. Achieve AI-Adaptive Speech Recognition to Adapt to Different Scenarios and User Habits: The AI-adaptive speech recognition module adopts a three-layer model architecture, dynamically adjusting the recognition threshold and feature weights by combining noise feature labels. This allows for targeted adaptation to speech recognition needs in different noise scenarios. Simultaneously, the BIO slot labeling model accurately extracts core information such as adjustment actions, objects, and amplitudes, ensuring the accuracy of ordinary seat control operations. Furthermore, by constructing a user seat control preference model, combined with cosine similarity matching and fuzzy inference algorithms, fuzzy command disambiguation is achieved. A reliability assessment and minimum clarification interaction mechanism are also implemented to further improve the accuracy of command parsing, adapting to different users' adjustment habits, reducing invalid operations, and making seat adjustments more tailored to user needs. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a schematic diagram of a module for an intelligent seat control system based on AI voice and Star Flash technology.

[0019] Figure 2 This is a flowchart illustrating the speech acquisition and noise reduction process.

[0020] Figure 3 This is a flowchart illustrating the AI ​​speech recognition and ambiguous intent disambiguation process.

[0021] Figure 4 This is a flowchart for the Level 1 safety interlock control process.

[0022] Figure 5 This is a flowchart for the Level 2 safety interlock control process. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] like Figure 1 As shown, an intelligent seat control system based on AI voice and Star Flash technology is disclosed. The system includes: a voice acquisition and processing module, an AI adaptive voice recognition module, a Star Flash wireless communication module, a seat control execution module, and a safety interlock module. The voice acquisition and processing module is used to acquire the voice signals of the target occupants in a directional manner, and to perform noise reduction processing on the voice signals of the target occupants according to the improved adaptive noise reduction algorithm to obtain voice segments with noise feature labels.

[0026] It should be noted that the application scenarios of this invention cover typical scenarios such as idling, low-speed congestion, high-speed driving, air conditioning on, audio playback, and conversation among multiple passengers.

[0027] like Figure 2 As shown in this embodiment, the specific process of obtaining speech segments with noise feature labels includes: Since noise sources in the vehicle environment are dispersed, including conversations from the driver, front passenger, and rear passengers, as well as the operating sounds of in-vehicle equipment, the voice acquisition and processing module uses a 4-microphone array, which is embedded on both sides of the seat headrest and in front and rear of the vehicle ceiling to form a 360° sound field coverage, adapting to different voice interaction scenarios for different passengers, such as the driver adjusting the seat or the front passenger controlling the seat angle by voice. During the data acquisition process, the beamforming algorithm parameters are dynamically adjusted based on the in-vehicle scenario: When the vehicle is idling, the engine noise is approximately 50-60dB, mainly in the low-frequency range of 250-500Hz. The beamwidth corresponding to this parameter is automatically adjusted to 30°, with a focus on the driver's seat and passenger area to suppress ambient noise in the rear. When the vehicle is traveling at high speeds (typically ≥80km / h), wind noise increases significantly to approximately 70-85dB, primarily propagating through door seals and thin steel plates. The beamwidth is automatically narrowed to 15° to enhance the directional capture of the target occupant's voice source. Spatial filtering technology is used to suppress external traffic noise, including but not limited to the roar of large trucks and the impact of gravel on the road. Noise from non-target areas, including but not limited to conversations among rear passengers and children crying, is also filtered to ensure that only the voice signals of the target occupants are collected, avoiding the mixing of irrelevant noise.

[0028] It should be noted that the microphone array in this embodiment has a voice activity detection function. When the voice of the target occupant is detected, it automatically starts to collect data. When there is no voice, it enters a low-power standby state to reduce the collection of invalid noise and reduce the subsequent processing pressure on the vehicle system.

[0029] In this embodiment, different types of noise will generate noise echoes inside the vehicle. In the vehicle environment, noise echo interference includes, but is not limited to, three types of noise echo scenarios. A reference signal needs to be set according to the actual scenario, and adaptive echo cancellation is performed using a normalized least mean square algorithm. The echo suppression weights are iteratively updated using an iteration step size to ensure rapid tracking of echo changes. The first type of noise echo scenario is engine noise echo. When the vehicle is idling or driving at low speed, engine noise enters the vehicle through the front fender, superimposing with the occupants' voices to form an echo. The engine ECU's output speed signal is used as an auxiliary reference, combined with the engine noise collected by the microphone, to construct a precise reference signal, focusing on suppressing low-frequency echoes in the 250-500Hz range. The second type of noise echo... The first scenario involves wind noise echoes. At high speeds, the wind noise entering the vehicle from the front is reflected by the windows and doors, creating echoes. In this case, the wind speed signal collected by the vehicle's wind speed sensor is combined with the wind noise signal collected by the microphone as a reference signal to specifically suppress high-frequency wind noise echoes in the 1000-2000Hz range. The second scenario involves audio system echoes. When music or navigation voice is played in the car, the sound is reflected by the ceiling and seat surfaces and collected by the microphone, creating echoes. In this case, the audio system output signal is directly used as a reference signal, and the echo suppression weights are dynamically adjusted based on changes in audio volume. This prevents damage to the target occupant's voice signal during echo cancellation and also prevents voice interruption due to over-suppression. To address the time-varying nature of in-vehicle echo paths, such as changes in passenger numbers or seat positions altering the in-vehicle sound field, the echo suppression weights are updated every 100ms to ensure stable echo cancellation performance and adapt to dynamic changes in different in-vehicle scenarios.

[0030] It should be noted that the speech signal after echo cancellation will still have a small amount of environmental noise remaining. Secondary noise reduction is required based on segmented spectrum subtraction and noise estimation algorithms to match the non-stationary characteristics of vehicle noise. In this embodiment, the echo-cancelled speech signal is processed by frame segmentation, with a frame length of 20ms and a frame shift of 10ms, to adapt to the real-time requirements of in-vehicle voice control. For example, the occupant's voice control of the seat needs to respond quickly with a delay of ≤100ms. A minimum tracking noise estimation algorithm is used to track the noise power spectrum of each frame in real time and dynamically adjust the noise estimation strategy for different in-vehicle scenarios. For example, at idle speed, the noise is mainly stable low-frequency engine noise, so a leading silence estimation is used, taking the first 0.5 seconds of the speech as the noise sample. At high speed, the noise is non-stationary wind noise and tire noise, so dynamic minimum tracking is used, updating the noise estimation value every 5 frames to avoid noise estimation deviation. The framed speech signal is denoised by segmented spectral subtraction, with a lower spectral limit parameter α=0.01 to prevent excessive noise reduction and retain the key features of the target occupant's speech, including but not limited to tone of voice and command keywords, resulting in clear speech segments after denoising.

[0031] During noise reduction processing, noise feature parameters corresponding to the current speech segment are calculated in real time and precisely quantified to match the actual noise characteristics of the vehicle, including but not limited to the signal-to-noise ratio (SNR). Reference thresholds are set based on different vehicle scenarios, with a target SNR of ≥25dB ​​at idle and ≥18dB at high speed, calculated using a posterior SNR calculation method. Real-time calculation, in which The power spectrum of the current frame. The noise power spectrum is estimated to accurately reflect the intensity ratio of speech to noise. The noise power spectrum is converted to the frequency domain through short-time Fourier transform, capturing the frequency distribution characteristics of the noise in real time. For example, the noise power spectrum is concentrated in 250-500Hz at idle, 1000-2000Hz at high speed, and 500-1000Hz when the air conditioner is on. The noise interference type is determined by five typical noise labels preset in the vehicle scenario: engine noise, wind noise, audio echo, tire noise, and multi-occupant interference noise. The current noise interference type is automatically identified by the frequency distribution and fluctuation amplitude of the noise power spectrum. For example, if the noise power spectrum is concentrated in 250-500Hz and the fluctuation is small, it is determined to be engine noise.

[0032] The real-time calculated signal-to-noise ratio (SNR), noise power spectrum, and noise interference type are encoded using a binary encoding method. The noise interference type occupies 4 bits, corresponding to 5 typical noise types and reserving 1 unknown noise type. The SNR occupies 8 bits with a quantization range of 0-40dB. The noise power spectrum occupies 16 bits with a quantization frequency range of 20-20000Hz. This encoding results in a 28-bit noise feature tag, ensuring the tag information is concise and includes key noise parameters. This noise feature tag is then bound to the denoised speech segment using a timestamp synchronization mechanism to ensure a one-to-one correspondence between the tag and the speech segment, avoiding misalignment.

[0033] It should be noted that scene identifiers are embedded in the noise feature labels, which are provided as a reference by the AI ​​adaptive speech recognition module. When the AI ​​adaptive speech recognition module receives a speech segment with noise feature labels, it can adaptively adjust the recognition model parameters according to the noise type and signal-to-noise ratio in the labels to improve the accuracy of speech recognition and adapt to the complex and ever-changing noise environment in the vehicle.

[0034] The AI ​​adaptive speech recognition module is used to perform AI adaptive processing on the speech segments with noise feature labels to obtain ordinary parsed seat control transactions; and to perform fuzzy intent disambiguation processing on the ordinary parsed seat control transactions to obtain the corresponding disambiguated parsed seat control transactions.

[0035] It should be noted that the AI ​​adaptive speech recognition module adopts a three-layer model architecture, including an acoustic model, a language model, and a BIO slot labeling model. The three layers work together to adapt to complex in-vehicle scenarios, such as high-speed wind noise, idling engine noise, interference from multiple occupants talking, and audio playback, while taking into account both the real-time performance and accuracy of speech recognition.

[0036] To further explain, the BIO slot labeling model is based on the BIO labeling system, which defines slot boundaries through Begin-Inside-Outside labels. It combines a BiLSTM model (Bidirectional Long Short-Term Memory) or a BERT-BiLSTM model (Bidirectional EncoderRepresentations from Transformers-Bidirectional Long Short-Term Memory) to explicitly model label dependencies, reduce illegal label sequences, and significantly improve text ambiguity handling capabilities by utilizing pre-trained semantic knowledge.

[0037] like Figure 3 As shown, in this embodiment, the specific process of obtaining a normal parsed seat control transaction includes: Speech segments with noise feature labels are input into an acoustic model. This model uses an improved CNN-LSTM architecture, which is suitable for optimizing non-stationary noise in vehicles. Based on the noise feature labels bound to the speech segments, the recognition threshold and feature weights are dynamically adjusted to achieve adaptive noise scene. Specifically, the acoustic model first extracts key information from the noise feature labels, namely the noise interference type, signal-to-noise ratio (SNR), and noise power spectrum. Then, it adjusts the model parameters. For example, when the label is identified as high-speed wind noise or high-frequency non-stationary noise with an SNR ≤ 18dB, it automatically increases the extraction of high-frequency speech features, such as the weight of keywords corresponding to occupant voice commands, such as "adjust," "backrest," or "raise," while reducing the recognition sensitivity of high-frequency noise bands to avoid misidentification of keywords caused by wind noise interference. For another example, when the label is identified as idling engine noise or low-frequency stationary noise with an SNR ≥ 25dB, it focuses on strengthening the extraction of low-frequency speech features, suppressing the 250-500Hz engine noise band, lowering the recognition threshold by 5%, and adapting to the driver's soft commands when idling, such as whispering "move the seat back" when the driver is focused on the road. When the label is identified as multi-occupant interference noise or random frequency, it activates the speech separation algorithm, combining the spatial positioning information of the microphone array to separate the target occupant's speech from the interference speech, and then performs feature extraction to avoid rear-seat occupant conversations triggering seat control commands.

[0038] It's worth noting that the acoustic model incorporates a dedicated corpus for in-vehicle voice commands, covering commonly used seat control phrases in various scenarios. For example, at high speeds, a driver might say, "Raise the seat a little to see the road ahead," or in traffic jams, "Recline the seat to rest." The model is trained specifically to address the characteristics of in-vehicle voice commands, including faster speech, concise tone, and potential regional accents. This ensures that noise-robust speech segments are accurately converted into text data, avoiding transcription errors caused by the specific characteristics of in-vehicle scenarios, such as mistakenly transcribing "recline back" as "backrest upright." Furthermore, the acoustic model employs a lightweight design, adapting to the hardware resources of the in-vehicle system to avoid excessive memory and computing power consumption. The transcription latency is controlled within 30ms, allowing ample time for subsequent processing.

[0039] To further explain, the main body of the acoustic model adopts a hybrid architecture of CNN and LSTM. The front-end feature extraction layer, as the entry point of the acoustic model, is responsible for receiving speech segments with noise feature labels. It first preprocesses the speech signal and connects it with the frame-segmentation processing of the speech acquisition module, using the parameters of 20ms frame length and 10ms frame shift to ensure temporal consistency. Then, it extracts 13-20 dimensional key speech features through Mel frequency cepstral coefficients to mimic the characteristics of human hearing, focusing on capturing the core frequencies of passenger speech in the vehicle scenario, filtering redundant frequency components, adapting to the lightweight computing power requirements of the vehicle system, and avoiding invalid features occupying resources. The CNN convolutional layer consists of three layers to suppress local noise in the in-vehicle environment. It uses 3×3 convolutional kernels for feature filtering, and each convolutional layer is followed by batch normalization and ReLU activation to reduce gradient vanishing and preserve the integrity of speech features. The first convolutional layer focuses on suppressing low-frequency noise (corresponding to idling engine noise), the second layer focuses on suppressing mid-frequency noise (corresponding to air conditioning and tire noise), and the third layer focuses on enhancing high-frequency speech features (corresponding to occupant command keywords), achieving frequency-band noise suppression and feature enhancement. The LSTM temporal modeling layer consists of two layers to adapt to the temporal characteristics of in-vehicle speech (e.g., occupant commands are spoken quickly and concisely). It is responsible for capturing the temporal dependencies of the speech signal, avoiding transcription errors caused by incomplete speech segments, such as mistakenly transcribing "seat move backward" as "seat move forward." An attention mechanism is embedded to focus on keywords in the speech commands, such as "adjust," "backrest," and "raise," reducing the interference weight of non-target noise and improving the accuracy of temporal feature extraction. Backend output layer: Employs fully connected layers and the Softmax activation function to convert extracted speech features into text probability distributions and output the most fitting text data; and performs lightweight processing on the fully connected layers to reduce parameter redundancy and control latency to within 30ms, adapting to the ≤100ms response requirements of in-vehicle seat control.

[0040] In this embodiment, a language model is used for structured parsing, tailored to the in-vehicle seat control scenario. Text data output from the acoustic model is input into the language model, which is built based on semantic rules specific to in-vehicle seat control. This model abandons general semantic parsing logic, focusing on adapting to the core needs of seat control. It utilizes statistical rules and lightweight semantic encoding—that is, adopts a lightweight Transformer architecture—to reduce computational consumption and achieve rapid structured parsing of text data. Specifically, the language model first preprocesses the text data, filtering out invalid and redundant information, such as common interjections like "ah," "oh," and "um," while retaining the core semantics. Then, through statistical rules, it matches commonly used sentence structures for in-vehicle seat control, including "action + object" or "object + action + amplitude," such as adjusting the seat back or raising the seat by 5 centimeters. The text data is then broken down to distinguish between core instructions and redundant information.

[0041] Based on actual in-vehicle interaction scenarios, the language model is optimized for different occupants' command habits: driver commands are mostly concise and direct, emphasizing quick response; commands from front passenger or rear passenger need to be more detailed and require accurate object recognition; and fuzzy commands are adapted by first parsing the basic control direction to lay the groundwork for subsequent fuzzy intent disambiguation.

[0042] It should be noted that the language model also has scene adaptation capabilities, combining scene identifiers in noise feature labels to optimize parsing logic; for example, when driving at high speed, it prioritizes parsing commands related to driving safety, such as raising the seat and straightening the backrest, while ignoring commands that may affect safety, such as reclining the seat; when idling or parked, it can parsing various adjustment commands normally, improving the flexibility of interaction.

[0043] In this embodiment, the text data, after being structured and parsed by the language model, is input into the BIO slot annotation model. This model is specifically designed to annotate four core slots for in-vehicle seat control scenarios: adjustment actions, such as raising, lowering, moving forward, reclining, and folding down; adjustment objects, such as driver's seat, front passenger seat, rear seats, seat back, and seat cushion; adjustment range, such as 1 cm, 5 cm, slight adjustment, and full adjustment; and target modes, such as comfort mode, driving mode, and rest mode. Through slot annotation, the core control information in the text is accurately extracted. After annotation, legality verification and format standardization are performed to conform to the actual hardware limitations and safety requirements of in-vehicle seat control, avoiding invalid or non-compliant commands.

[0044] It's important to note that the legality verification focuses on two aspects. First, the legality of the adjustment slot itself. For example, the adjustment range must be within the seat's hardware travel range. If the seat's fore-and-aft adjustment range is 0-20 cm, and the range is labeled "30 cm," it's considered invalid, the slot is removed, and other valid information is retained. Second, the legality of the scenario. Combining the scenario identifier with the noise feature label, the verification checks whether the instruction matches the current scenario. For example, if "seat recline" is labeled during high-speed driving, it's considered a scenario violation, and the instruction is temporarily stored. Format standardization integrates the extracted slot information according to a fixed format: adjustment object + adjustment action + adjustment range + target mode, unifying the expression. For example, lowering the passenger seat slightly is standardized as "Adjustment object: passenger seat; Adjustment action: lower; Adjustment range: small; Target mode: none." Similarly, raising the driver's seat by 5 cm and switching to comfort mode is standardized as "Adjustment object: driver's seat; Adjustment action: raise; Adjustment range: 5 cm; Target mode: comfort mode."

[0045] Through the collaborative processing of the above three-layer model, a clear, standardized, and in-vehicle-compatible general parsing seat control transaction is finally obtained, ensuring the accuracy of subsequent fuzzy intent disambiguation and command execution, while taking into account the real-time performance and security of in-vehicle interaction, adapting to different noise scenarios and different occupants' command habits, and truly meeting the actual needs of in-vehicle use.

[0046] It should be noted that, in combination with actual in-vehicle interaction scenarios, for common fuzzy commands in ordinary seat control transactions, such as adjusting the seat to a more comfortable position, backrest to a more reclined position, preference modeling is used to achieve accurate disambiguation, ensuring that the disambiguation results match the usage habits of the target occupant, while also taking into account the real-time nature and convenience of in-vehicle interaction.

[0047] In this embodiment, the specific process of obtaining the corresponding disambiguation parsing seat control transaction includes: Collect and organize historical adjustment data of target occupants, clarify the scope and type of data collection, and ensure the authenticity and relevance of the samples. The historical adjustment data includes, but is not limited to, historical adjustment records, collecting seat adjustment operation records of target occupants within the past 3 months, including but not limited to voice command adjustment, manual button adjustment, etc., and associating them with scene identifiers, such as idling, high speed, congestion, and parking for rest. Habitual amplitude data, statistically analyze the target occupants' preferences for small, medium, and large amplitude adjustments, and combine this with the vehicle seat hardware travel, i.e., the maximum fore-aft adjustment travel of 0-20 cm and the backrest angle adjustment range of 90°-160°, to standardize the fuzzy amplitudes. For example, a driver's slight backward movement corresponds to an actual adjustment of 3-5 cm, a medium backward movement corresponds to an increase in backrest angle of 10-15°, and a front passenger's slight lowering corresponds to a decrease in seat height of 2-3 cm. Preferred angle data, record the target occupants' commonly used seat angles in different scenarios, such as a driver's commonly used backrest angle of 85° and seat cushion tilt angle of 3° during commuting, and a commonly used backrest angle of 100° and seat cushion tilt angle of 5° after a 1-hour long-distance drive.

[0048] Weakly supervised samples are constructed based on historical data to form training and testing sets. Sample labels are generated through scene association and operation frequency statistics. To address the diversity of in-vehicle scenarios, samples from different noise environments and driving states are supplemented, avoiding model bias caused by a single sample. A lightweight logistic regression and collaborative filtering algorithm is used to construct a seat control preference model for the target occupant.

[0049] It should be noted that the construction process of the seat control preference model is as follows: The Transformer encoder was set to 3 layers, the attention heads to 4, and the hidden layer dimension to 64 to avoid excessive model complexity that could strain the vehicle's computing power. The Adam optimizer was used, with initial learning rate and decay coefficients set, 100 iterations, and a batch size of 32 to accommodate the vehicle's memory limitations. A hybrid loss function was employed, combining mean squared error loss and cross-entropy loss to balance regression accuracy of the adjustment magnitude with classification accuracy of the adjustment action and object. The formula is as follows: ;in, This is used to adjust the mean squared error loss of the amplitude quantization value, reflecting the deviation between the model's predicted amplitude and the historical actual amplitude; This is a quantified value of the actual historical amplitude. The magnitude quantization value for the model prediction, where N is the number of training samples; It is used to adjust the cross-entropy loss of objects, actions, and target patterns, reflecting the classification accuracy; (This is a one-hot encoding for the tag, where 1 represents the corresponding category and 0 represents other). Predict the probability of the corresponding category for the model; The total number of categories is 12, consisting of 4 categories of objects, 5 categories of actions, and 3 categories of patterns. , The weighting coefficients are set to 0.6 and 0.4 respectively to prioritize the accuracy of the adjustment range and to match passenger preferences.

[0050] After each iteration, the model accuracy and mean squared error are calculated using the validation set. Training is stopped when the validation set accuracy is ≥90% and the mean squared error is ≤0.8. If the target is not reached after 100 iterations, an early stopping strategy is adopted to avoid overfitting. The model is then tested based on the test set to ensure its generalization ability. The model is considered to be successfully trained if the test set accuracy is ≥90% and the mean squared error is ≤1.0.

[0051] In this embodiment, based on a successfully trained seat control preference model, fuzzy instructions in ordinary parsing seat control transactions are disambiguated. A dual algorithm of cosine similarity matching and fuzzy inference is used to ensure the accuracy of disambiguation. The specific process is as follows: The fuzzy instructions in ordinary seat control transactions are divided into two categories: single fuzzy instructions and compound fuzzy instructions.

[0052] Calculate the cosine similarity between the semantic vector of the fuzzy instruction and the semantic vector of historical samples in the seat control preference model, and select the bottom 5 samples based on similarity ranking as the disambiguation candidate set. The similarity calculation formula is as follows: ;in, The semantic vector representing the current fuzzy instruction; Semantic vectors of historical samples; The value represents the similarity score, ranging from [0, 1]. The closer the value is to 1, the higher the semantic consistency between the instruction and the historical samples.

[0053] Based on the candidate set samples and the current in-vehicle scenario, fuzzy inference is performed to determine the optimal disambiguation result. The core logic is as follows: The scene matching score between each candidate sample in the disambiguation candidate set and the current scene is calculated. Scenes with identical scenes are set to a matching score of 1, while scenes with similar scenes are set to an undefined score. Scene matching degree with large differences is set to For scenarios with completely different settings, the matching degree is set to... .

[0054] Calculate the overall score for each candidate sample. ;in This represents the similarity between the corresponding candidate samples; This represents the scene matching degree of each candidate sample; This indicates that the satisfaction scores of the candidate samples have been normalized to A satisfaction rating of 5 points corresponds to 1.0, a satisfaction rating of 4 points corresponds to 0.8, a satisfaction rating of 3 points corresponds to 0.6, a satisfaction rating of 2 points corresponds to 0.4, a satisfaction rating of 1 point corresponds to 0.2, and a satisfaction rating of 0 points corresponds to 0. The adjustment parameter corresponding to the candidate sample with the lowest comprehensive score is selected as the preliminary disambiguation result of the fuzzy instruction.

[0055] In this embodiment, a confidence assessment model is used to score the initial disambiguation results to determine their reliability and avoid erroneous adjustment. The specific process and formula are as follows: A lightweight logistic regression model is used, with the similarity, scene matching, and satisfaction scores corresponding to the preliminary disambiguation results as inputs, and the confidence score as output. The range is [0,1]. The training process is synchronized with the seat control preference model, using the same batch of weakly supervised samples to ensure the accuracy of the evaluation.

[0056] The confidence score is calculated using the following formula: ;in This represents the Sigmoid activation function, used to map the output to the [0,1] interval. , , These represent feature weights, obtained through model training, with optimal values ​​of 0.5, 0.25, and 0.25 respectively. This is the bias term, and its value after training is -0.3.

[0057] Set confidence threshold ; when When the disambiguation result is highly reliable, the preliminary disambiguation result is directly organized into a standard format of "adjustment object + adjustment action + adjustment range + target mode" to obtain the disambiguation analysis of the seat control transaction and send it to the Star Flash wireless communication module. when If the disambiguation result is unreliable, a minimal clarification interaction is triggered to avoid erroneous adjustments. The minimal clarification interaction uses a simplified voice query, only asking questions about ambiguous points, without increasing the operator's workload. For example, if the ambiguous command is "adjust the seat to be more comfortable", the minimal clarification question would be "Do you want to adjust the seat back or the seat cushion?" After the operator replies via voice, the reply information is added to the normal seat control transaction, and the "cosine similarity matching + fuzzy inference + confidence assessment" process is re-executed until the corresponding confidence score is greater than or equal to the confidence threshold, thus obtaining the final disambiguation resolution of the seat control transaction. If the operator does not reply within 3 seconds, the command is temporarily stored, and the query is asked again after an interval of 5 seconds. If there is no reply after two consecutive queries, the command is canceled to avoid consuming vehicle system resources.

[0058] It should be noted that the accuracy of the disambiguation results is calculated every 10 days. When the accuracy is below 90%, the confidence assessment model is incrementally trained to ensure the accuracy of the assessment and adapt to the dynamic changes in passenger preferences.

[0059] The StarScan wireless communication module is used to classify and process the disambiguation parsing seat control transactions and generate corresponding control transaction instruction sequences.

[0060] It should be noted that the StarSpark wireless communication module is a low-latency, highly reliable, and interference-resistant short-range wireless communication unit built on the StarSpark protocol. It primarily undertakes core functions such as wireless transmission of seat control commands, priority scheduling, multi-device linkage, and safety interlock signal forwarding. It is a key transmission hub for realizing intelligent in-vehicle seats and interconnected cabin ecosystems. It ensures no delay, no packet loss, and no false triggering of commands even in scenarios with complex in-vehicle noise, concurrent devices, and high-speed driving. It also supports collaborative work with multiple nodes such as in-vehicle systems, mobile phones, smart home devices, and in-vehicle sensors, providing occupants with a safe, convenient, and seamless intelligent seat control experience. For example, in leisure-mode scenarios such as parking and waiting at coffee shops or shopping mall entrances, when a vehicle is temporarily parked at the entrance of a coffee shop, shopping mall, or service area, and passengers need to wait and rest inside the car, they can trigger leisure mode via voice or a single button. The StarSpark wireless communication module receives and forwards control commands with low latency and high speed; the seat automatically performs actions such as reclining, leg rest support, and lumbar support relaxation; it can also simultaneously link the air conditioning, ambient lighting, and audio system via the StarSpark wireless communication module to quickly enter a comfortable resting state. The entire process is free from the delays and interference of traditional wiring harnesses, achieving one-click comfort in in-car waiting scenarios. For example, in convenient exit scenarios when arriving at home or parking, when the vehicle arrives at the residential garage or the destination is ready to exit, the vehicle's engine is turned off and the door opening signal is transmitted in real time via the StarSpark wireless communication module; the StarSpark wireless communication module can immediately issue seat control commands to automatically move the seat backward, increasing legroom and entry / exit space; this makes it easier for the elderly, children, passengers wearing high heels, or carrying items to exit the vehicle, improving convenience. Thanks to the rapid response and highly reliable transmission of the StarSpark wireless communication module, seat movements and door opening actions are synchronized, enabling seamless and convenient exit from the vehicle. In scenarios with independent voice control for multiple occupants, where multiple people simultaneously use seat adjustment functions, the StarSpark wireless communication module supports concurrent transmission across multiple nodes and seats. The driver, front passenger, and rear passengers can each use voice control independently without interference. Even at high speeds, in windy conditions, or in noisy environments, commands are transmitted stably without loss of control or accidental touches. In remote pre-adjustment and welcome scenarios, when occupants approach the vehicle or use remote control, the StarSpark wireless communication module wirelessly connects to terminals such as mobile phones and smartwatches. For example, it can pre-activate seat heating or ventilation, or automatically move back to welcome passengers when unlocking the vehicle, ensuring comfort upon entry.

[0061] In this embodiment, the process of generating the corresponding control transaction instruction sequence includes: A low-latency, short-range wireless link within the vehicle is established based on the NearLink protocol, and corresponding node pairing and link synchronization are configured. Seat control transactions are categorized into high-priority and normal-priority commands based on disambiguation analysis of their security level and urgency. QoS is configured for high-priority commands, and a resource unit dynamic scheduling algorithm is used to prioritize the allocation of time-frequency resources and transmission time slots. A sequence number, checksum, and retransmission flag are added to each command, and an acknowledgment-retransmission mechanism is constructed to correct abnormal commands. The commands are then sorted according to priority level and time order to form a corresponding control transaction command sequence.

[0062] It should be noted that QoS is a service quality (QoS) policy system that explicitly guarantees end-to-end low-latency transmission of high-priority traffic, reduces the risk of critical service interruption caused by network congestion, and significantly improves the predictability of service quality for critical business applications by utilizing pre-configured policy rules.

[0063] In this embodiment, the specific process of constructing multi-scenario security interlocking rules and multi-level anomaly degradation strategies includes: Based on the vehicle safety knowledge graph, the correlation strength between seat safety status and various driving scenarios is calculated. By traversing the network and accumulating the contribution scores of multi-source seat safety status parameters to different driving scenarios, several candidate driving scenarios are selected. The correlation strength calculation formula is as follows: ,in Indicates the seat safety status With driving scenarios The correlation strength is denoted as [0,1], and the closer the value is to 1, the higher the correlation. The total number of seat safety status parameters includes seat adjustment range, motor speed, locking status, tilt sensor data, etc., in this embodiment. ; For the first The weights of each safety state parameter are determined using the Analytic Hierarchy Process (AHP), where the weight of the seat locking state is... Adjusting travel weights Motor speed weight Tilt angle parameter weights Voltage parameter weights Temperature parameter weighting ; Seat safety status Next The actual values ​​of each parameter; For driving scenarios Next Standard thresholds for each parameter; For the first The maximum value of each parameter is used for normalization to ensure that the calculation results are within the normal range. Interval.

[0064] Set the correlation strength threshold ,when At that time, the driving scenario Candidate driving scenarios were included, and idling, low-speed congestion, high-speed driving, temporary stopping, and parking were selected to cover mainstream in-vehicle usage scenarios.

[0065] Retrieve historical driving scenarios and seat control records, extract historical seat safety status features of similar driving scenarios, calculate the parameter similarity between the seat safety status parameters of the current candidate driving scenario and the historical seat safety status features of the corresponding driving scenario, and iterate through all candidate driving scenarios according to their contribution scores. The parameter similarity calculation formula is as follows: , in The parameter similarity value ranges from [0,1]. The closer the value is to 1, the better the current parameter matches the historical features. For the current candidate scenario, the first Real-time values ​​of several safety status parameters; For the first time in a similar historical scenario The average value of each safety status parameter; For the first The minimum value of each parameter; Consistent with the above correlation strength calculation, ensure that the parameter weights are consistent.

[0066] Set confirmation threshold ,when When the current scenario does not belong to the candidate driving scenario, other candidate scenarios are re-matched; When determining the corresponding candidate driving scenarios, the safety interlock constraint type is as follows: High-speed driving scenario, i.e., constraint type A: seat reclining is prohibited, backrest tilt angle ≤110°, adjustment speed ≤5mm / s; Low-speed congestion scenario, i.e., constraint type B: small adjustments are allowed, backrest tilt angle ≤130°, adjustment speed ≤8mm / s; Idle scenario, i.e., constraint type C: regular adjustments are allowed, backrest tilt angle ≤140°, adjustment speed ≤10mm / s; Temporary parking scenario, i.e., constraint type D: most adjustments are allowed, backrest tilt angle ≤150°, seat not allowed to be completely moved out of the guide rail; Parking scenario, i.e., constraint type E: no excessive constraints, only limiting the adjustment range within the hardware limits.

[0067] By combining the seat control entities associated with the corresponding safety interlock constraint types, the specific adjustment mechanisms corresponding to the target seat state parameters, such as seat motors, locking solenoid valves, tilt adjusters, etc., and the safety interlock adjustment action types, such as limiting, speed limiting, locking, prohibition, etc., are matched to clarify the executable safety restriction rules. For example, in the high-speed driving scenario, the rules corresponding to constraint type A are as follows: the upper limit of seat back tilt adjustment is 110°, exceeding which triggers the tilt limit command; the seat reclining action is directly prohibited, triggering the locking command; the speed of all adjustment actions is ≤5mm / s, exceeding which triggers the speed limit command; if the tilt angle exceeds the threshold during adjustment, the adjustment is immediately stopped and rolled back to 110°. The rules of all candidate scenarios are integrated to form multi-scenario safety interlock rules. The rules are encoded in XML format, which facilitates quick parsing by the StarFlash module and the execution module.

[0068] The system collects the operational status and fault codes between modules in real time, calculates the correlation weight between abnormal indicators of each module and the system fault level, sets anomaly judgment criteria based on historical anomaly cases, and then generates corresponding anomaly degradation strategies. The abnormal indicators include star-flash communication delay, execution module response time, safety interlock module detection frequency, and seat motor current deviation. The correlation weight calculation formula is as follows: ,in For the a-th abnormal indicator and the a-th The association weight of level faults, with a value range of [0,1]; The first in historical cases The first abnormal indicator triggered the first The number of level-one failures; For the first The total number of times each abnormal indicator triggers all fault levels; Abnormal indicators such as star-flash communication delay, execution module response time, safety interlock module detection frequency, and seat motor current deviation are marked. Mark the star flash communication delay; Mark the response time of the execution module; Mark the detection frequency of the safety interlock module; Mark the deviation of the seat motor current.

[0069] In this example, the system fault levels are divided into three levels: Level 1 fault, which is a minor anomaly, such as slightly high communication delay; Level 2 fault, which is a moderate anomaly, such as execution module response timeout; and Level 3 fault, which is a severe anomaly, such as abnormal motor current or lock failure. Based on historical anomaly sample data, the correlation weight between each anomaly indicator and the fault level is obtained. The anomaly judgment criteria are set according to the 3σ principle, including Level 1 anomaly indicators, Level 2 anomaly indicators, or Level 3 anomaly indicators.

[0070] Based on the associated weights, corresponding anomaly degradation strategies are selected; for each anomaly degradation strategy, control schemes are formulated to maintain the current position, roll back to a safe posture, or switch local control, and corresponding trigger conditions and execution flows are bound to form corresponding multi-level anomaly degradation strategies, specifically including: The first-level anomaly degradation strategy is triggered when any first-level anomaly indicator is triggered and there are no second- or third-level anomalies. The control scheme is to maintain the current seat position. The execution process is to pause the current adjustment command, maintain the current state of the seat, monitor the anomaly indicator in real time, and update it every 500ms. If the anomaly persists for 10 seconds without relief, it is upgraded to the second-level anomaly degradation strategy. Level 2 anomaly downgrade strategy, trigger condition: any level 2 anomaly indicator is triggered, or level 1 anomaly persists for 10 seconds without relief: the control scheme is to roll back to the safe posture, the execution process is: immediately stop the adjustment command, roll the seat back to the safe reference position in the current scenario, the rollback speed is ≤3mm / s, after the rollback is completed, lock the seat to prevent further adjustment, and send an anomaly prompt to the vehicle system; Level 3 anomaly downgrade strategy, trigger condition: any Level 3 anomaly indicator is triggered, or a Level 2 anomaly persists for 5 seconds without relief: the control scheme is to switch to local control and emergency lock, execution process: immediately disconnect the star flash wireless communication link, switch to the seat local button control mode, and at the same time trigger the seat emergency lock, lock all adjustment mechanisms, prohibit any adjustment operation, send an emergency fault alarm to the vehicle system, and simultaneously record fault codes and abnormal data for subsequent inspection.

[0071] The seat control execution module is used to execute the control transaction instruction sequence and collect feedback data in real time during the execution of the control transaction instruction sequence.

[0072] The safety interlock module is used to construct multi-scenario safety interlock rules and multi-level anomaly degradation strategies; based on the multi-scenario safety interlock rules and real-time driving status data, it performs first-level safety interlock control on the feedback data; based on the multi-level anomaly degradation strategies, it performs second-level safety interlock control on the seat; based on the two-level safety interlock control, the safety of seat adjustment is improved.

[0073] like Figure 4 As shown, in this embodiment, the specific process of performing first-level safety interlock control on the feedback data based on the multi-scenario safety interlock rules and real-time driving status data includes: Real-time driving status data is collected, including vehicle speed, engine speed, parking status, and seatbelt wearing status, and the data is normalized. The correlation strength between the current real-time driving status data and the corresponding feedback data for the primary safety interlock control is calculated. The correlation strength calculation formula is as follows: ;in The strength of the first-level safety interlock control association is defined, with a value range of [0,1]. This is the normalized value for vehicle speed; This is the normalized value for engine speed; In park status (0 or 1); This represents the seatbelt wearing status (0 or 1). It should be noted that vehicle speed has the highest weight in this formula because it is the core factor affecting seat adjustment safety, with the greatest difference in safety constraints between high and low speeds.

[0074] Set the threshold for the first-level safety interlock control association strength. , , ,according to The value of the filter is used to select the safety interlock rules for the current candidate driving scenario: when When the scenario is determined to be a high-speed driving scenario, the safety interlock rule of constraint type A is matched. when If the scenario is determined to be a low-speed congestion or idling scenario, further distinction is made based on the engine speed. According to the engine speed, rules of constraint type C and constraint type B are matched respectively. when : If the scenario is determined to be a temporary parking scenario, the rule of constraint type D will be matched; when : If the scenario is determined to be a parking scenario, then the rule of constraint type E is matched.

[0075] In this embodiment, based on the seat adjustment object and actuator associated with the safety constraint rules, a limiting instruction corresponding to the feedback data is matched; the limiting instruction includes a seat motor voltage limiting instruction, a seat motor current limiting instruction, an adjustment direction limiting instruction, and an adjustment stroke parameter limiting instruction. The parameters of each limiting instruction are calculated as follows: Regarding the motor voltage limiting command: ,in This is the voltage limit threshold. The rated voltage of the seat motor is used in this embodiment. ; For the level 1 safety interlock control association strength, The larger the value, the lower the voltage limit threshold, which reduces the motor speed and achieves speed control. For motor current limiting commands: ,in This is the rate limiting threshold. The rated current of the seat motor is specified in this embodiment. ; The larger the value, the stricter the current limiting, to avoid motor overload during high-speed adjustment; For adjustment direction limiting commands: Set the upper limit of adjustment in each direction according to the matched constraint type. For example, in a high-speed scenario (constraint type A), the upper limit of the backrest tilt angle. The calculation formula is: 160° is the maximum reclining angle of the backrest. The larger the value, the lower the upper limit. For the instruction to adjust the stroke parameter limit: ,in This is the travel limit threshold. In this embodiment, the maximum travel of the seat's fore-and-aft adjustment is... This ensures a more conservative seat adjustment range in high-speed scenarios, avoiding any impact on driving safety.

[0076] The aforementioned adjustment limit command is sent to the seat control execution module. The execution module adjusts the motor operating parameters according to the command and provides real-time feedback on the execution status. The safety interlock module compares the feedback data with the adjustment limit command threshold. If the feedback data exceeds the threshold, a stop command is immediately triggered to achieve first-level safety interlock control and output the constraint status.

[0077] like Figure 5 As shown, in this embodiment, the specific process of performing secondary safety interlock control on the seat according to the multi-level anomaly degradation strategy includes: The system monitors the operational status of each module during the execution of control transaction command sequences in real time, including star-flash communication delay, execution module response time, safety interlock module detection frequency, and seat motor current deviation, and standardizes abnormal indicators.

[0078] Calculate the correlation strength between the corresponding module's abnormal indicators and the secondary safety interlock control of each level of abnormal degradation strategy. The correlation strength calculation formula is as follows: ,in The strength of the secondary safety interlock control association for the level b anomaly degradation strategy, i.e. This corresponds to the first, second, and third level anomaly degradation strategies; For the first The absolute value of each abnormal indicator after standardization.

[0079] Based on the association strength of the secondary safety interlock control, the association weight is set, and the anomaly degradation strategy with the highest adaptability is selected. The selection rule is: calculate R2(1). , The value of R2(b) is used to select the corresponding degradation strategy with the largest value as the adaptation strategy; if Maximum and If no obvious abnormality is found, the downgrade strategy will not be triggered; if Maximum and This triggers the secondary exception degradation strategy; if Maximum and This triggers a three-level exception degradation strategy; if Maximum and This triggers the first-level exception degradation strategy.

[0080] Based on the control mode and safety posture associated with the aforementioned abnormal degradation strategy, the corresponding control scheme is executed to achieve secondary safety interlock control of the current seat operating state. The specific execution process and parameter calculation are as follows: The first-level exception degradation strategy is triggered, the current seat position maintenance plan is executed, and the duration of the maintained state is calculated. , The larger the size, the longer it lasts, but the maximum duration is no more than 20 seconds; The baseline maintenance time is used; during the maintenance period, abnormal indicators are collected every 500ms, and real-time calculations are performed. ,like Release the hold state and restore normal regulation; if Immediately upgrade to a Level 3 anomaly degradation strategy; if the duration reaches [a certain threshold]... The strategy has been upgraded to a Level 2 anomaly downgrade strategy.

[0081] Triggering the level 2 anomaly degradation strategy, executing the rollback safety posture plan, and calculating the rollback target position: backrest tilt angle. ,in The safety reference tilt angle for the current scenario, such as a high-speed scenario. Idle scenario (etc.); front and back positions ,in In this embodiment, the seat is in the middle position for fore-and-aft adjustment. Rollback speed , The baseline rollback speed; The larger the value, the slower the rollback speed, ensuring a smooth rollback process. During rollback, the seat position feedback data is monitored in real time. When the deviation between the feedback position and the target position is ≤0.5cm, rollback stops and the seat lock is triggered. Lockback time is specified. , Based on the baseline locking time, an abnormal notification is sent to the vehicle's infotainment system: "Seat malfunction, has been rolled back to a safe position."

[0082] Triggering a level-three abnormal degradation strategy, the system executes a switch to local control and emergency locking mechanism, first disconnecting the StarFlash wireless communication link and switching the delay. After the switch is complete, voice control and remote control via the Star Flash feature will be disabled, leaving only local seat button control, and only emergency adjustments, such as straightening the backrest, will be allowed; at the same time, an emergency lock will be triggered, locking the current. Ensure reliable locking; after locking, monitor the motor current in real time. Immediately disconnect the motor power supply, send an emergency fault alarm to the vehicle's infotainment system, and record the fault code simultaneously in the format of abnormality type-abnormality level-time stamp, such as "motor current abnormality-level three-20260416100000", to facilitate troubleshooting by maintenance personnel.

[0083] It should be noted that during the secondary safety interlock control process, the control effect evaluation value is calculated in real time. , The value range is [0,1]. The closer to 1, the better the control effect; when If this occurs, it indicates that the current degradation strategy is not adaptable enough. Immediately switch to a higher-level degradation strategy to ensure the safe operation of the seat and form a closed-loop control.

[0084] Optionally, in this embodiment, those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0085] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0086] If the integrated units in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in the aforementioned computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause one or more electronic devices to execute all or part of the steps of the methods described in the various embodiments of this application.

[0087] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0088] In the several embodiments provided in this application, it should be understood that the disclosed application can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of units or modules may be electrical or other forms.

[0089] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0090] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0091] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. An intelligent seat control system based on AI voice and star-flash technology, characterized in that, The system includes: a voice acquisition and processing module, an AI adaptive voice recognition module, a star-flash wireless communication module, a seat control execution module, and a safety interlock module; The voice acquisition and processing module is used to acquire the voice signal of the target occupant in a directional manner, and to perform noise reduction processing on the voice signal of the target occupant according to the improved adaptive noise reduction algorithm to obtain voice segments with noise feature labels. The AI ​​adaptive speech recognition module is used to perform AI adaptive processing on the speech segments with noise feature labels to obtain ordinary parsed seat control transactions; and to perform fuzzy intent disambiguation processing on the ordinary parsed seat control transactions to obtain the corresponding disambiguated parsed seat control transactions. The StarScan wireless communication module is used to classify and process the disambiguation parsing seat control transactions and generate corresponding control transaction instruction sequences. The seat control execution module is used to execute the control transaction instruction sequence and collect feedback data in real time during the execution of the control transaction instruction sequence; The safety interlock module is used to construct multi-scenario safety interlock rules and multi-level anomaly degradation strategies; based on the multi-scenario safety interlock rules and real-time driving status data, it performs first-level safety interlock control on the feedback data; based on the multi-level anomaly degradation strategies, it performs second-level safety interlock control on the seat; based on the two-level safety interlock control, the safety of seat adjustment is improved.

2. The intelligent seat control system based on AI voice and star-flash technology according to claim 1, characterized in that, The process of obtaining speech segments with noise feature labels includes: A microphone array is used to spatially directionally acquire occupant speech. Beamforming algorithm is used to lock the direction of the target sound source and suppress environmental noise in non-target areas. Engine noise echo, wind noise echo, and audio system echo are used as reference signals. Adaptive echo cancellation is performed using a normalized least mean square algorithm, and the echo suppression weights are iteratively updated. Based on segmented spectral subtraction and noise estimation algorithms, the target occupant speech signal after adaptive echo cancellation is denoised to obtain the corresponding speech segments. The corresponding signal-to-noise ratio, noise power spectrum, and noise interference type are calculated in real time. The corresponding signal-to-noise ratio, noise power spectrum, and noise interference type are encoded as noise feature tags and bound to the denoised speech segments for output, forming speech segments with noise feature tags.

3. The intelligent seat control system based on AI voice and star-flash technology according to claim 2, characterized in that, The process of obtaining a normal parsed seat control transaction includes: A three-layer model architecture is adopted, including an acoustic model, a language model, and a BIO slot labeling model. Speech segments with noise feature labels are input into the acoustic model, and the recognition threshold and feature weights are dynamically adjusted according to the noise feature labels to convert noise-robust speech segments into text data. The text data is input into the language model, and the text is structured and parsed using statistical rules and lightweight semantic coding. Then, the corresponding adjustment actions, adjustment objects, adjustment amplitudes, and target modes are extracted according to the BIO slot labeling model, and legality verification and format standardization are performed to obtain ordinary parsed seat control transactions.

4. The intelligent seat control system based on AI voice and star-flash technology according to claim 3, characterized in that, The process of obtaining the corresponding disambiguation resolution seat control transaction includes: Weakly supervised samples are constructed based on the target occupant's historical adjustment records, habit range, and preference angle data, thereby building a seat control preference model for the target occupant. Based on this seat control preference model, fuzzy instructions for ordinary parsing seat control transactions are disambiguated. A confidence score for the corresponding disambiguation result is calculated based on a confidence assessment model. When the confidence score is greater than or equal to a confidence threshold, the corresponding disambiguated parsing seat control transaction is directly obtained. When the confidence score is less than the confidence threshold, a minimum clarification interaction is triggered, thereby obtaining the corresponding disambiguated parsing seat control transaction.

5. The intelligent seat control system based on AI voice and star-flash technology according to claim 4, characterized in that, The process of generating the corresponding control transaction instruction sequence includes: A low-latency, short-range wireless link within the vehicle is established based on the NearLink protocol, and corresponding node pairing and link synchronization are configured. Seat control transactions are categorized into high-priority and normal-priority commands based on disambiguation analysis of their security level and urgency. QoS is configured for high-priority commands, and a resource unit dynamic scheduling algorithm is used to prioritize the allocation of time-frequency resources and transmission time slots. A sequence number, checksum, and retransmission flag are added to each command, and an acknowledgment-retransmission mechanism is constructed to correct abnormal commands. The commands are then sorted according to priority level and time order to form a corresponding control transaction command sequence.

6. The intelligent seat control system based on AI voice and star-flash technology according to claim 5, characterized in that, The process of constructing multi-scenario security interlocking rules includes: Based on the vehicle safety knowledge graph, the correlation strength between seat safety status and various driving scenarios is calculated. By traversing the network and accumulating the contribution scores of multi-source seat safety status parameters to different driving scenarios, several candidate driving scenarios are selected. Historical driving scenarios and seat control records are retrieved, and historical seat safety status features of similar driving scenarios are extracted. The similarity between the seat safety status parameters of the current candidate driving scenario and the historical seat safety status features of the corresponding driving scenario is calculated. All candidate driving scenarios are traversed according to their contribution scores. If the parameter similarity is less than the confirmation threshold, the scenario is not considered a candidate driving scenario. If the parameter similarity is greater than or equal to the confirmation threshold, the safety interlock constraint type of the corresponding candidate driving scenario is determined. The seat control entity associated with the corresponding safety interlock constraint type is combined with the specific adjustment mechanism and safety interlock adjustment action type corresponding to the target seat status parameters to clarify the executable safety restriction rules, thereby forming multi-scenario safety interlock rules.

7. The intelligent seat control system based on AI voice and star-flash technology according to claim 6, characterized in that, The process of constructing a multi-level anomaly degradation strategy includes: The system collects the operating status and fault codes between modules in real time, calculates the correlation weight between abnormal indicators of each module and the system fault level, sets abnormal judgment criteria through historical abnormal cases, and generates corresponding abnormal degradation strategies. Based on the correlation weight, the system selects the corresponding abnormal degradation strategies. For each abnormal degradation strategy, the system formulates control schemes to maintain the current position, roll back to a safe posture, or switch local control, and binds corresponding trigger conditions and execution processes to form corresponding multi-level abnormal degradation strategies.

8. The intelligent seat control system based on AI voice and star-flash technology according to claim 7, characterized in that, Based on the multi-scenario safety interlock rules and real-time driving status data, the process of performing first-level safety interlock control on the feedback data includes: The system calculates the correlation strength between the current real-time driving status data and the corresponding feedback data for the primary safety interlock control. Based on the correlation strength, it selects safety interlock rules for the current candidate driving scenario. According to the seat adjustment object and actuator associated with the safety constraint rules, it matches the adjustment limitation instructions in the corresponding feedback data. The adjustment limitation instructions include seat motor voltage limitation instructions, seat motor current limitation instructions, adjustment direction limitation instructions, and adjustment stroke parameter limitation instructions, thereby realizing primary safety interlock control and outputting the constraint status.

9. The intelligent seat control system based on AI voice and star-flash technology according to claim 8, characterized in that, The process of implementing secondary safety interlock control for the seat according to the multi-level anomaly degradation strategy includes: The system monitors the operating status of each module during the execution of control transaction instruction sequences, calculates the correlation strength between the corresponding module's abnormal indicators and the secondary safety interlock control of each level of abnormal degradation strategy, sets correlation weights based on the correlation strength of the secondary safety interlock control, and selects the abnormal degradation strategy with the highest adaptability. Based on the control mode and safety posture associated with the abnormal degradation strategy, the system executes the corresponding control scheme, thereby realizing the secondary safety interlock control of the current seat operating status.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which is executed by a processor to implement the intelligent seat control system based on AI voice and star-flash technology as described in any one of claims 1-9.

Images