Intelligent shock absorption and noise reduction control method for rail transit

Abstract

The invention discloses an intelligent shock absorption and noise reduction control method for rail transit, and belongs to the technical field of rail transit noise and vibration control. The method aims at solving the problems that traditional passive control measures are poor in adaptive capacity and limited in frequency band, and effective prospective offset is difficult to achieve due to system delay and inaccurate reference signals in existing active control. The method comprises the following steps: synchronously acquiring vibration and noise signals through a sensor array; calculating the contribution weight of vibration to noise by using a cross-domain modal correlation module based on an attention mechanism, and fusing the features; predicting an evolution sequence of vibration and noise in a specific time period in the future by adopting a long-short-term memory network model; synthesizing a target reference signal according to the prediction sequence, and generating an anti-phase control signal through an adaptive filtering algorithm; finally, it is ensured that the total delay of the system is smaller than the feedforward advance time, and an actuator is driven to achieve accurate foresight counteracting. The method is mainly used for improving the self-adaptive capacity and the control effect of the rail transit damping and noise reduction system.

Description

Technical Field

[0001] This invention belongs to the field of rail transit noise and vibration control technology, specifically relating to an intelligent vibration reduction and noise control method for rail transit. Background Technology

[0002] During rail transit operation, the vibrations and noise generated when trains pass over track structures have become key factors affecting the comfort of the environment and the safety of equipment along the line. For a long time, passive control technologies, such as installing elastic sleepers, vibration damping pads, or sound barriers, have been widely adopted in engineering practice. These methods rely on the physical properties of the structure itself to dissipate or isolate energy. While they have the advantages of simple structure and high reliability, their vibration reduction and noise reduction performance has gradually shown limitations as train speeds and loads continue to increase.

[0003] The main problem with passive control measures lies in their fixed effect and limited frequency band. Their performance largely depends on pre-designed structural parameters, and once installed, they are difficult to dynamically adjust according to changes in external excitation conditions. For example, their optimal vibration reduction frequency band is often narrow, making it difficult to effectively suppress broadband vibrations and noise generated by a combination of factors such as wheel-rail relationships and track irregularities across the entire frequency range. When train speed, axle load, or track conditions change, existing passive measures may not always operate at their optimal state, leading to a weakening of their control effect on specific frequency bands. The root cause of this problem is that passive control systems lack the ability to sense and respond to time-varying excitations.

[0004] To overcome the shortcomings of passive control, active control technology has been explored and researched. Its basic principle is to achieve interference cancellation by applying a secondary cancellation source that is out of phase with the target vibration or noise. However, applying the concept of active control to the complex scenario of rail transportation faces many unique challenges. First, vibration and sound sources exhibit significant time-varying and uncertainties. Train operation is a dynamic process; the amplitude, frequency components, and spatial distribution of the vibrations and noise it generates evolve rapidly with changes in speed, position, and track structure. This requires the control system to possess extremely high response speed and predictive capability; otherwise, accurate cancellation cannot be achieved. Traditional feedback control strategies, due to system delays, always lag behind disturbances in their control actions, performing poorly when dealing with high-frequency components and even posing a risk of instability.

[0005] If a feedforward control strategy is adopted, its effectiveness is highly dependent on the accurate acquisition of the reference signal and the precise prediction of future disturbances. In the rail transit environment, the generation mechanisms of vibration and noise are complex, and they belong to two different physical domains: structure-based sound and airborne sound. The coupling relationship between them is difficult to describe accurately using simple physical models. Previous solutions have mostly relied on single-type sensor signals or simplified theoretical models, which are unable to fully capture the complex paths and coupling mechanisms of vibration and sound transmission, resulting in low-quality reference signals. Without high-quality reference signals, feedforward control cannot generate accurate reverse control signals. In addition, to achieve effective global control, a large-scale array of sensors and actuators needs to be deployed over a large spatial area. This brings about the challenges of real-time synchronization, transmission, and processing of massive amounts of data, placing almost stringent demands on the computing power, communication bandwidth, and algorithm efficiency of the control system.

[0006] In summary, existing technologies suffer from inherent drawbacks such as insufficient adaptive capabilities in complex time-varying environments, difficulty in accurately modeling cross-physical domain disturbance coupling relationships, and system delays that make it difficult to guarantee real-time control performance. These factors collectively constitute the technical obstacles to achieving efficient, reliable, and intelligent vibration and noise reduction in the rail transit field. Summary of the Invention

[0007] One object of the present invention is to solve at least the above-mentioned problems and / or defects, and to provide at least the advantages described below.

[0008] One objective of this invention is to address the problems in the field of rail transit, where the generation mechanisms of vibration and noise are complex and time-varying, traditional passive vibration reduction and noise reduction methods have poor adaptability and limited bandwidth, while existing active control technologies suffer from system delays leading to lagging control actions and difficulty in accurately acquiring reference signals to achieve effective forward-looking cancellation.

[0009] One objective of this invention is to address the problem that the huge computational load caused by the cross-domain modal correlation module in calculating the correlation between all vibration channels and noise makes it difficult to meet the hard real-time requirements of the control system.

[0010] One objective of this invention is to address the inherent problem of error accumulation in long-term forecasting, as well as the issue of decreased forecast accuracy due to model mismatch or environmental disturbances, which in turn affects the accuracy of feedforward control.

[0011] One objective of this invention is to address the problem that the uncertainty and jitter of the total system delay time Td make it difficult to stably meet the core timing constraint that it must be less than the feedforward control lead time, thus leading to the failure of forward control.

[0012] One objective of this invention is to address the problem that adaptive filtering algorithms suffer from low computational efficiency due to the traditional time-domain point-by-point sampling processing method, making it difficult to meet the real-time requirements for generating the entire control signal sequence.

[0013] One objective of this invention is to solve the problem that the actual output of the actuator cannot accurately reproduce the ideal control signal generated by the algorithm due to the inherent nonlinear characteristics of the actuator, thereby reducing the final performance of the entire active control system.

[0014] One objective of this invention is to address the problem that methods for pre-screening vibration channels based on static prior knowledge cannot adapt to the dynamically changing environment of the track system, and that lightweight attention mechanisms may result in a loss of model accuracy.

[0015] One objective of this invention is to address the problem that staged prediction strategies rely excessively on the accuracy of the endpoint prediction, and that their online error compensation methods may introduce instability due to lag and simple coupling.

[0016] One objective of this invention is to address the problem that using injected test signals for delay calibration interferes with the normal operation of the system, and that a single calibration result cannot reflect the dynamic changes in system delay.

[0017] One objective of this invention is to address the inherent problem of large block delay in frequency domain block processing methods, and the impact of linear convolution errors introduced by cyclic convolution on the real-time performance and accuracy of control signals.

[0018] One object of the present invention is to provide an intelligent vibration reduction and noise control method for rail transit, the method comprising the following steps: Data acquisition steps: Simultaneously acquire vibration signals collected by multiple accelerometers deployed along the track and noise signals collected by microphones; Association and fusion steps: The vibration signal and noise signal are input into a cross-domain modal association module; the cross-domain modal association module calculates the contribution weight of each vibration signal channel to the noise signal features based on the attention mechanism, and performs weighted fusion of the vibration signal features according to the contribution weight, and outputs a fused cross-domain feature vector; Prediction steps: Input the cross-domain feature vector into a pre-trained long short-term memory network model. The model outputs a prediction sequence of the target vibration main frequency, amplitude and phase within the feedforward control lead time Ta, as well as a prediction sequence of the target point noise sound pressure level. Synthesis step: Based on the predicted sequence, synthesize a target reference signal sequence corresponding to the feedforward control lead time Ta; Generation step: Input the target reference signal sequence into an adaptive filtering algorithm to generate a control signal sequence with the same amplitude but opposite phase; Output steps: The control signal sequence is output to the actuator in real time, wherein the total system delay time Td from the start of calculation to the actuator's action is strictly less than the feedforward control lead time Ta, so as to achieve forward-looking cancellation of target vibration and noise.

[0019] Preferably, in the intelligent vibration reduction and noise reduction control method for rail transit, the execution mode of the cross-domain modal association module in the association and fusion step includes: Pre-screening sub-step: Based on prior knowledge of the track structure, select a subset of specific vibration channels from all vibration signal channels whose acoustic contribution to the target noise reduction point is higher than a preset threshold; Parallel computation sub-step: A lightweight multi-head attention mechanism is used to calculate the contribution weights between the signal and noise signal features of each channel in the specific vibration channel subset in parallel; Weighted fusion sub-step: Based on the calculated contribution weights, the signal features of the specific vibration channel subset are weighted and fused to generate the fused cross-domain feature vector.

[0020] Preferably, in the intelligent vibration reduction and noise control method for rail transit, in the prediction step, the long short-term memory network model makes predictions in the following manner: The first stage, coarse prediction, involves predicting the final amplitude of the dominant vibration frequency at the end of the feedforward control lead time Ta based on the input cross-domain feature vector. end Phase ƒ of final value end ; Second-stage sequence generation: using the final value amplitude... end Phase ƒ of final value end For strong constraints, a smooth transition sequence of vibration amplitude and phase from the current time to time Ta, which conforms to the final value constraint, is generated through a built-in decoder. Online compensation correction: The error between the predicted value at the current moment and the actual sensing measurement value at the current moment is obtained in real time, and the error is used as a feedback correction amount and injected into the generation process of the decoder to dynamically fine-tune the future prediction sequence that has not yet been output.

[0021] Preferably, in the intelligent vibration reduction and noise control method for rail transit, the method further includes a timing management step, which runs independently of the acquisition to output step and includes: System delay calibration: During system startup or idle periods, inject a test signal into the control loop, measure and record the total system delay Td from signal input to actuator response output. current ; Lead time adaptive adjustment: adjust the measured Td currentMultiplying this by a preset safety factor yields the current required feedforward control lead time Ta. required And dynamically adjust the prediction length in the prediction step to make it equal to Ta. required ; Buffer management: Before the output step, a first-in-first-out (FIFO) buffer is set up; the control signal sequence is temporarily stored in this buffer after generation and waits for a fixed buffer time T. buffer Then, it is sent to the actuator; the buffer time T buffer According to formula T buffer =Ta required -Td current The calculations show that jitter is used to eliminate computational delays.

[0022] Preferably, in the intelligent vibration reduction and noise reduction control method for rail transit, the adaptive filtering algorithm is executed in the generation step as follows: Block processing: The target reference signal sequence is divided into consecutive data blocks, and each data block is processed uniformly; Frequency domain transformation: Transforms the reference signal sequence of the current data block to the frequency domain using a fast Fourier transform; Frequency domain filtering: In the frequency domain, the transformed reference signal is multiplied by the coefficients of the current frequency domain filter to directly generate a control signal block in the frequency domain; Time-domain transformation: The frequency-domain control signal block is transformed back to the time domain through inverse fast Fourier transform to obtain the corresponding time-domain control signal sequence; Block update: After obtaining the actual error signal of a complete data block, the filter coefficients are updated in one block in the frequency domain.

[0023] Preferably, in the intelligent vibration reduction and noise reduction control method for rail transit, a feedforward compensation and protection step is added in the output step before sending the control signal sequence to the actuator: Inverse model feedforward compensation: The control signal sequence is input into an actuator inverse model compensator; the actuator inverse model is a mathematical inverse model that is pre-established by the system to describe the dynamic relationship between the ideal input and the actual output of the actuator; the compensator outputs a pre-distorted compensated control signal, which drives the actuator so that the actual output force or sound wave of the actuator can approach the requirements of the original control signal. Output amplitude limiting protection: The amplitude of the compensated control signal is monitored in real time. If the amplitude exceeds the saturation threshold of the linear operating region of the actuator, it is limited to within the saturation threshold to prevent the actuator from entering the nonlinear saturation region and to protect the equipment safety.

[0024] Preferably, in the intelligent vibration reduction and noise reduction control method for rail transit, the execution mode of the cross-domain modal association module in the association and fusion step includes: Online saliency assessment: For each vibration channel, the coherence function values ​​of its signal and the target noise signal in the time and frequency domains are calculated in real time to serve as the saliency index of that channel at the current moment; Dynamic channel competition: All channels are sorted according to their significance index, and the top N channels with the highest index are selected to form the current optimal vibration channel subset, where N is a preset fixed value. Attention weight calculation: An attention mechanism is used to calculate the contribution weight of the noise signal characteristics only for the channels in the current optimal vibration channel subset; Weighted fusion: Based on the calculated contribution weights, the signal features of the current optimal vibration channel subset are weighted and fused to generate the fused cross-domain feature vector.

[0025] Preferably, in the intelligent vibration reduction and noise reduction control method for rail transit, the prediction step further includes a dynamic prediction mode selection and compensation step, which includes: Confidence assessment: The Long Short-Term Memory network model outputs a high-resolution final value. end Phase ƒ of final value end At the same time, it outputs the confidence score C of its predicted value; Pattern fusion: A mixing factor α is set, the value of which is dynamically adjusted by the confidence score C; the decoder generates the final prediction sequence P. final The formula is: P final =α*P anchor +(1-α)*P recursive , where P anchor For a smooth transition sequence anchored at the final value, P recursive The sequence generated by the model's recursive prediction; Weighted compensation: When performing online error compensation, the feedback correction amount is weighted, and future time points that are similar to the current error frequency characteristics are given higher correction weights.

[0026] Preferably, in the intelligent vibration reduction and noise reduction control method for rail transit, the system delay calibration sub-step in the timing management step is executed in the following non-disruptive manner: Undisturbed Delay Estimation: Continuously monitor the control signal sequence and the actual effect signal fed back by sensors near the actuator; calculate the cross-correlation function between the control signal and the actual effect signal, find the time offset corresponding to its maximum peak value, and use this time offset as the current total system delay time Td. current The estimated value; Feedforward prediction: Establish a regression model of system load and computation latency, predict the trend of computation latency changes based on the current core utilization and memory usage of the processor, and then apply this model to the Td... current The estimated value is then fine-tuned using feedforward. Feedback calibration: Adjusting the feedforward Td current The buffer management sub-step is run, and the deviation between the actual control effect and the expected target is continuously compared; based on the deviation, a PID controller is used to calibrate the parameters of the regression model online.

[0027] Preferably, in the intelligent vibration reduction and noise reduction control method for rail transit, the adaptive filtering algorithm in the generation step is executed in the following optimized manner: Hybrid domain processing: The target reference signal sequence is divided into continuous data blocks; for each data block, it is further divided into a first half and a second half; the frequency domain block processing method is used to generate control signals for the second half of the data blocks; for the first half of the data blocks, a time domain filtering algorithm is used for parallel computation to significantly reduce the processing delay caused by waiting for the entire data block. Adaptive block length adjustment: Real-time monitoring of the main frequency of the target reference signal; dynamic selection of the optimal data block length based on the Nyquist sampling theorem and the available computing resources of the current system; when the main frequency is high, a smaller block length is automatically selected to reduce latency; when the main frequency is low, a larger block length is automatically selected to improve computational efficiency. Error compensation filtering: After the time domain transformation, a digital post-filter is applied to the generated time domain control signal sequence. The frequency response characteristics of the filter are designed to be opposite to the distortion characteristics introduced by circular convolution during the frequency domain block processing, in order to compensate for the aforementioned linear convolution error.

[0028] The present invention has at least the following beneficial effects: This invention constructs a complete "sensing-fusion-prediction-control" intelligent closed loop, comprehensively utilizing deep learning and adaptive control technologies to achieve proactive and precise cancellation of broadband, time-varying vibrations and noise in rail transit. Its beneficial effects lie in its ability to adapt to the train's operating state and overcome system delays by predicting future disturbances, thereby significantly improving the real-time performance and effectiveness of vibration and noise reduction, and overcoming the inherent defects of traditional methods such as response lag and narrow control bandwidth.

[0029] This invention significantly reduces the computational complexity of the cross-domain correlation module by introducing channel pre-screening based on prior knowledge and lightweight parallel attention computation. Its beneficial effect lies in its ability to efficiently complete the fusion processing of multi-source signals while meeting the system's hard real-time requirements, laying a solid foundation for subsequent accurate prediction and ensuring the speed of the entire control loop.

[0030] This invention effectively suppresses error accumulation in long-term time-series prediction by employing staged prediction combined with online error compensation. Its beneficial effects include improving the accuracy and robustness of the model's prediction of future vibration and noise sequences. Even in the face of model mismatch or environmental disturbances, it can be corrected through real-time feedback, thereby ensuring the quality of the feedforward control signal.

[0031] This invention introduces an independent timing management step to achieve online calibration of system delay and adaptive adjustment of feedforward lead. Its advantages lie in dynamically ensuring that the total system delay is always less than the core constraint of feedforward control lead, and eliminating delay jitter through buffer management, thereby guaranteeing the stable implementation of forward control and the accuracy of output timing.

[0032] This invention employs an adaptive filtering algorithm with frequency domain block processing to transform time-consuming time-domain convolution operations into efficient frequency-domain complex multiplication. Its advantages lie in significantly improving the generation efficiency of control signal sequences, meeting the requirements of real-time processing of massive amounts of data, and providing computational power assurance for generating high-quality feedforward control signals.

[0033] This invention effectively isolates the influence of the actuator's own nonlinearity and saturation characteristics on the control effect by introducing actuator inverse model feedforward compensation and output limiting protection. Its beneficial effect is that it ensures the actual output of the actuator accurately tracks the ideal command of the algorithm and stably constrains the system within the linear operating range, thereby significantly improving the final performance and reliability of the entire active control system.

[0034] This invention achieves adaptive screening of vibration channels by employing online saliency assessment and a dynamic channel competition mechanism. Its advantages lie in its ability to automatically focus on the most critical vibration source based on real-time data, eliminating reliance on static prior knowledge and thus improving the system's adaptability and robustness under different operating environments while maintaining computational efficiency.

[0035] This invention enhances the intelligent decision-making capability of the prediction process by introducing confidence assessment and multi-mode fusion prediction, and employing a weighted error compensation strategy. Its beneficial effects include reducing reliance on a single prediction mode, improving the algorithm's fault tolerance under complex conditions, and enhancing the effectiveness and stability of compensation through more refined error allocation, thereby obtaining a more reliable prediction sequence.

[0036] This invention achieves intelligent management of system delay by employing a perturbation-free delay estimation and a feedforward-feedback composite timing control strategy. Its advantages lie in its ability to accurately estimate delay online without injecting interference test signals, predict delay change trends for feedforward compensation, and ensure long-term accuracy through a feedback loop, thereby constructing a higher-order, more intelligent timing control system.

[0037] This invention deeply optimizes the frequency domain block processing algorithm by employing hybrid domain processing, adaptive block length, and error-compensated filtering strategies. Its advantages include significantly reducing inherent block latency, eliminating linear convolution errors, and intelligently selecting optimal parameters based on signal characteristics. Thus, while inheriting the high efficiency of frequency domain computation, it achieves lower latency and higher signal fidelity.

[0038] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Detailed Implementation

[0039] The present invention will now be described in further detail so that those skilled in the art can implement it based on the description.

[0040] This invention provides an intelligent vibration reduction and noise control method for rail transit. The method performs the following steps: Acquisition step: Simultaneously acquiring vibration signals collected by multiple accelerometers deployed along the track and noise signals collected by microphones; Association and fusion step: Inputting the vibration signals and noise signals into a cross-domain modal association module; The cross-domain modal association module, based on an attention mechanism, calculates the contribution weight of each vibration signal channel to the noise signal features, and performs weighted fusion of the vibration signal features according to the contribution weight, outputting a fused cross-domain feature vector; Prediction step: Inputting the cross-domain feature vector into a pre-trained long short-term memory network model, which outputs a... The system comprises: a predicted sequence of the target vibration dominant frequency, amplitude, and phase within a feedforward control lead time Ta; a predicted sequence of the target point noise sound pressure level; a synthesis step: synthesizing a target reference signal sequence corresponding to the feedforward control lead time Ta based on the predicted sequences; a generation step: inputting the target reference signal sequence into an adaptive filtering algorithm to generate a control signal sequence with the same amplitude but opposite phase; and an output step: outputting the control signal sequence to the actuator in real time, wherein the total system delay time Td from the start of calculation to the actuator's action is strictly less than the feedforward control lead time Ta, so as to achieve forward-looking cancellation of target vibration and noise. The control signal sequence refers to a series of discrete voltage or digital command values ​​arranged chronologically to drive the actuator.

[0041] To implement the method described in this invention, a complete hardware system is constructed, with the following core components and requirements: Sensor Array: Employing industrial-grade MEMS accelerometers (frequency response range at least 0.5Hz-5kHz) and condenser microphones (frequency response range 20Hz-20kHz). All sensors are hardware-synchronized via the IEEE 1588 (PTP) precision clock protocol to ensure a time synchronization error of less than 100 microseconds for full-domain data acquisition. Sensors are deployed linearly or in a matrix at key locations such as track fasteners, track bed, and tunnel walls, with spacing determined based on the vibration wave wavelength (typically 5-20 meters) to ensure effective capture of the spatial modes of vibration and sound propagation. Central Processing Unit: Employing an embedded processor with AI acceleration capabilities (such as NVIDIA Jetson AGX Orin, TI TDA4VM, etc.) or an industrial-grade IPC. It must possess at least TOPS-level integer and floating-point computing power to support real-time inference of LSTM models and attention mechanisms. Memory capacity should be no less than 8GB to ensure smooth frequency domain processing of large data blocks. Actuator Array: Depending on the control objective, select electric, electromagnetic, or piezoelectric actuators (for force cancellation) or speaker arrays (for sound cancellation). Actuators should have sufficient frequency response bandwidth (covering the target control frequency band) and output dynamic range. Their placement needs to be determined based on acoustic transfer function (ATF) experiments or simulations to form an effective canceling sound or force field. Communication System: Data transmission between sensors and the processing unit uses deterministic Ethernet (such as TSN) or high-speed fieldbus (such as EtherCAT) to meet the low latency and high reliability requirements for synchronous transmission of massive amounts of data.

[0042] The specific implementation process of the intelligent vibration reduction and noise control method for rail transit of the present invention is as follows.

[0043] First, a large number of acceleration vibration sensors and microphone noise sensors are deployed at key locations along the track and tunnel walls where vibration and noise reduction are required, forming a dense sensor array. These sensors require high-precision time synchronization to ensure that they can synchronously collect multi-dimensional vibration and noise signals generated when trains pass by.

[0044] The acquired multi-source signals are transmitted to the central processing unit in real time. At the core of this unit is a deep learning model pre-trained on a large amount of data. This model first runs a module called cross-domain modal correlation, which, based on the attention mechanism, intelligently analyzes the intrinsic correlation between each vibration sensor signal and the noise signal, and calculates the contribution of each vibration signal to the final noise, i.e., its weight. Subsequently, based on these weights, the features of all vibration signals are weighted and fused to form a fused feature vector that comprehensively reflects the vibration-acoustic coupling relationship.

[0045] Next, the fused feature vector is input into a Long Short-Term Memory (LSTM) network model. The core task of this network model is to perform look-ahead prediction. Based on the current fused features, it accurately predicts how the frequency, amplitude, and phase of the main vibrational components will evolve over a specific time period in the future, while also predicting the trend of noise sound pressure level changes at the target measurement point. The predicted output is a continuous sequence, not the value at a single time point.

[0046] Then, based on the predicted vibration frequency, amplitude, and phase sequence, the system synthesizes a sinusoidal reference signal sequence of the same length, representing the future ideal cancellation target. This reference signal sequence is fed into an adaptive filtering algorithm. Through iterative optimization, the algorithm quickly generates a control signal sequence with the same amplitude as the reference signal sequence but with exactly opposite phase.

[0047] Finally, the entire system ensures that the total time from signal calculation to the actuator's final action is shorter than the previously predicted future time. Once this condition is met, the generated control signal sequence is sent in real-time to the actuator array arranged on the track structure. Based on the received signals, the actuators output a canceling force or canceling sound wave that is out of phase with the predicted vibration and noise, thus effectively canceling them before the target vibration and noise actually arrive, achieving intelligent, proactive vibration and noise reduction.

[0048] Existing active vibration and noise reduction control systems based on feedback principles also collect vibration or noise signals through sensors, which are then processed by a controller to drive actuators to produce a cancellation effect. Compared to this existing technology, the essential difference and innovation of this invention are: First, the control principle is fundamentally different. Existing technologies use feedback control, the core feature of which is sensing the error that has already occurred and then generating control actions in response to the error. This is a reactive, lagging control mode, and its control effect is limited by system delay, performing poorly and easily becoming unstable when dealing with high-frequency disturbances. This invention, however, adopts an innovative paradigm combining feedforward control and predictive control. Its core is to predict future disturbances in advance and issue control commands ahead of time, a proactive control mode that fundamentally overcomes the bottleneck caused by system delay. Second, the signal processing methods are significantly different. Existing technologies typically process single vibration or noise signals or simply superimpose signals, failing to deeply explore the deep coupling relationship between vibration and noise signals from two different physical domains. This invention creatively introduces a cross-domain modal correlation module based on an attention mechanism, which can intelligently quantify the contribution of vibrations at different locations to noise, achieving deep fusion and understanding of multi-source heterogeneous signals, and providing a high-quality data foundation for accurate prediction. Third, the level of intelligence represents a generational leap. Existing controller algorithms are mostly traditional filtering or proportional-integral-differential algorithms, which cannot cope with the complex, time-varying nonlinear problems in rail transit scenarios. The core driving force of this invention is a trained long short-term memory network deep learning model, which enables it to accurately predict complex vibration and noise sequences—something traditional methods cannot achieve, representing a leap from automation to intelligence in control systems. In summary, this invention fundamentally differs from and significantly improves upon the closest existing technologies in terms of control principles, core signal processing algorithms, and system intelligence, thereby achieving superior vibration reduction and noise reduction performance.

[0049] In a preferred embodiment, in the intelligent vibration reduction and noise control method for rail transit, the cross-domain modal association module in the association and fusion step is executed in the following manner: a pre-screening sub-step: based on prior knowledge of the track structure, a subset of specific vibration channels with an acoustic contribution correlation higher than a preset threshold is selected from all vibration signal channels; a parallel computing sub-step: a lightweight multi-head attention mechanism is used to calculate the contribution weights between the signal and noise signal features of each channel in the specific vibration channel subset in parallel; and a weighted fusion sub-step: based on the calculated contribution weights, the signal features of the specific vibration channel subset are weighted and fused to generate the fused cross-domain feature vector.

[0050] It is important to note that the specific implementation of the cross-domain modal correlation module can adopt one of the following two preferred implementation methods, depending on the needs of the actual application scenario and the amount of available prior knowledge. This method is suitable for scenarios where the line is fixed, the operating conditions are relatively stable, and sufficient prior knowledge has been accumulated through previous simulations or tests. Its advantages are low computational load and high stability.

[0051] During system deployment, engineers first establish a priori knowledge base about the track structure based on historical data, simulation models, and acoustic theories. This knowledge base clarifies which vibration sensor locations, under different sections and operating conditions, have the most closely correlated signals with the noise at the target noise reduction point. During system operation, before the cross-domain modal correlation module starts calculations, it calls upon this priori knowledge base to quickly select a specific subset with the highest correlation from hundreds of vibration channels based on information such as the current train position and speed—for example, selecting the twenty most critical channels.

[0052] Subsequently, instead of calculating all channel data, the system only performs calculations on the selected twenty channels using a lightweight multi-head attention network (LMHA). This network has a simplified structure, with the following parameters: 4 attention heads and 64 dimensions for both keys and values. The total number of parameters can be kept below 10K, far lower than the standard Transformer model, thus meeting the computational constraints of the embedded platform. During computation, the LMHA module can be embedded and deployed on the NPU module of the aforementioned embedded processor. Utilizing the processor's multi-core parallel computing architecture, the weight calculation tasks for different channels are distributed to different cores to achieve parallel acceleration and meet the system's hard real-time requirements.

[0053] Finally, based on the calculated channel weights, the module performs a weighted summation of the signal features from these selected channels to generate a highly condensed, information-rich, fused cross-domain feature vector, which is then output to the subsequent prediction module. This entire process, through pre-screening and parallel computation, significantly reduces the scale and time of data processing.

[0054] The closest existing technology is a fusion method that directly calculates attention weights for all sensor channel signals. This method indiscriminately inputs all vibration and noise signals into a standard-sized attention network model, attempting to forcibly extract the correlations between all channels through the model's complex computational power. Compared with this existing technology, the essential differences and innovations of this invention are: First, the data processing strategy is fundamentally different. Existing technologies employ a "coarse-grained" data processing model, where the computational load increases quadratically with the number of sensor channels, facing significant real-time challenges and making it difficult to deploy in practical engineering. This invention proposes a "lean" two-stage processing strategy. By introducing a pre-screening step based on prior knowledge, it first focuses on the most critical subset from a massive number of channels, greatly reducing the input dimensionality of subsequent calculations and solving the computational bottleneck problem at its source. Second, the allocation of computing resources differs significantly. Existing technologies use powerful standard models to process large amounts of redundant data, resulting in low resource utilization. This invention uses a lightweight model optimized for specific tasks to process the selected high-value data, achieving precise and efficient utilization of computing resources. Meanwhile, the parallel computing architecture further unlocks the hardware's potential, pursuing ultimate processing speed while ensuring accuracy. Third, engineering feasibility and reliability are enhanced. Existing technologies, due to their excessive computational load, struggle to meet the hard real-time requirements of rail transit control systems, making engineering implementation difficult. Furthermore, their models are large and difficult to run on resource-constrained edge computing devices. This invention, through pre-screening and model lightweighting, enables complex deep learning algorithms to run stably and in real-time on practical hardware platforms, greatly improving the engineering value and reliability of the solution.

[0055] In summary, this invention, through its core design of pre-screening and parallel lightweight computation, successfully solves the engineering challenge of high computational complexity while inheriting the excellent feature extraction capabilities of deep learning, thus achieving a leap from theoretical algorithms to engineering applications.

[0056] In a preferred embodiment, in the intelligent vibration reduction and noise control method for rail transit, in the prediction step, the long short-term memory network model performs prediction as follows: First stage coarse prediction: Based on the input cross-domain feature vector, predict the final amplitude of the dominant vibration frequency at the end of the feedforward control lead time Ta. end Phase ƒ of final value end Second-stage sequence generation: using the final value amplitude... end Phase ƒ of final value endFor strong constraints, a built-in decoder generates a smooth transition sequence of vibration amplitude and phase from the current time to time Ta that conforms to the final value constraint; online compensation correction: the error between the predicted value at the current time and the actual sensing measurement value at the current time is obtained in real time, and the error is used as a feedback correction amount and injected into the generation process of the decoder to dynamically fine-tune the future prediction sequence that has not yet been output.

[0057] After receiving the fused cross-domain feature vectors, the Long Short-Term Memory (LSTM) network model performs its prediction process in two stages. In the first stage, the model does not directly predict the entire time series, but instead focuses on the final state, outputting the final amplitude and phase of the target oscillation dominant frequency at the end of the feedforward control lead time. These two final values ​​serve as anchor points for the predicted sequence, providing global constraints and objectives for the entire prediction.

[0058] In the second stage, a dedicated decoder within the model begins operation. This decoder uses the final value amplitude and phase predicted in the first stage as strong constraints to generate a complete sequence of vibration amplitude and phase from the current time to the target time. This sequence is not generated arbitrarily, but rather is a smooth transition curve converging to the final value, consistent with the inertial characteristics of the physical system.

[0059] Throughout the prediction process, the system also implements online compensation and correction. The model acquires the predicted value at the current moment in real time and compares it with the actual value measured by the sensor at that moment to calculate the prediction error. This error value is immediately used as a feedback correction amount and injected into the decoder's generation process to dynamically fine-tune the future prediction sequence that has not yet been output, enabling the prediction to correct deviations in a timely manner and adapt to environmental changes. During online compensation and correction, the error between the predicted value at the current moment and the actual sensor measurement value at that moment is acquired in real time. Specifically, to prevent instantaneous strong interference or abnormal sensor values ​​from contaminating the prediction sequence, the error needs to be smoothed by a first-order low-pass filter before injection, and when the absolute value of the error exceeds a preset safety threshold, the correction amount will be set to zero or significantly attenuated. The processed error is then used as a feedback correction amount and injected into the decoder's generation process to dynamically fine-tune the future prediction sequence that has not yet been output.

[0060] The Long Short-Term Memory (LSTM) network model incorporates an uncertainty quantification module to calculate the confidence score C (range 0-1) of its predictions. This score is calculated based on a combination of the model's internal state, the novelty of the input features, and recent prediction errors. The system sets a confidence threshold C. th (e.g., 0.6). When C ≥ C th When C is active, the system operates normally, employing the aforementioned two-stage prediction and online compensation strategy. <C thWhen this value is reached, it indicates low prediction reliability, and the system automatically activates a degraded prediction mode. Mode 1 (0.3≤C) <C th Mode 1: Abandoning complex final value prediction and sequence generation, it instead uses simplified linear extrapolation or AR models to directly predict vibration and noise for several future time steps based on recent historical data. Although accuracy decreases, it ensures output stability and real-time performance. Mode 2 (C<0.3): Determined as severely unreliable, feedforward predictive control is suspended, switching to pure feedback control or only outputting the basic damping signal, while simultaneously reporting a maintenance alarm requiring manual intervention. All mode transitions employ smooth transition algorithms to avoid abrupt changes in control commands.

[0061] The closest existing technology to this invention is a long-term time-series prediction method based on recursive prediction. This method uses predicted values ​​as input for the next prediction step, generating a prediction sequence for multiple future time steps through iterative iteration. Compared with this existing technology, the essential difference and innovation of this invention are: First, the prediction framework is fundamentally innovative. Existing technologies employ a simple recursive prediction mode, where prediction errors at each step are propagated and accumulated in all subsequent prediction steps, causing the predicted trajectory to gradually deviate from the true value, especially in long-term time-series prediction where error divergence is severe. This invention creatively adopts a two-stage prediction framework of endpoint anchoring plus sequence generation. First, the endpoint target is determined, and then a reasonable path to the target is generated, thus avoiding the accumulation and divergence of errors in the time dimension from a mechanism perspective. Second, the physical rationality of the prediction results is guaranteed. The prediction sequences generated by existing technologies are entirely data-driven and lack physical constraints, which may produce results that do not conform to physical laws. This invention, through endpoint constraints and smooth transition requirements, ensures that the generated prediction sequences not only satisfy data laws but also conform to the physical characteristics of the vibration system, improving the reliability and usability of the prediction results. Third, the system's adaptive capability is significantly enhanced. Existing technologies are purely open-loop predictions; once model mismatch or environmental disturbances occur, the entire prediction sequence will be biased. This invention introduces an online error compensation mechanism, forming a fast internal closed loop of prediction-measurement-correction. This enables the prediction system to learn and adapt online, greatly improving its robustness in changing environments.

[0062] In summary, this invention, through a two-stage prediction framework and an online compensation mechanism, solves the problem of error accumulation in long-term prediction while maintaining prediction accuracy, providing a more reliable and accurate prediction signal for feedforward control.

[0063] In a preferred embodiment, the intelligent vibration reduction and noise reduction control method for rail transit further includes a timing management step, which runs in parallel with the acquisition-to-output step and includes: system delay calibration: during system startup or idle periods, a test signal is injected into the control loop, and the total system delay time Td from signal input to actuator response output is measured and recorded. current ;Adaptive adjustment of lead time: adjust the measured Td current Multiplying this by a preset safety factor yields the current required feedforward control lead time Ta. required And dynamically adjust the prediction length in the prediction step to make it equal to Ta. required Buffer management: Before the output step, a first-in-first-out buffer is set up; the control signal sequence is temporarily stored in this buffer after generation and waits for a fixed buffer time T. buffer Then, it is sent to the actuator; the buffer time T buffer According to formula T buffer =Ta required -Td current The calculations show that jitter is used to eliminate computational delays.

[0064] The timing management steps operate in parallel with the main control loop. During system startup or operational downtime, this module injects a specific test signal into the control loop. The time interval between the input of this signal and the actuator producing a measurable response output is precisely measured and recorded as the current total system delay time.

[0065] Based on the measured system delay time, the module multiplies it by a safety factor greater than one to obtain the current required feedforward control lead time. The prediction module then dynamically adjusts the length of its prediction sequence according to this calculated lead time, ensuring that the prediction duration is always greater than the system delay.

[0066] In the signal output stage, the system employs a first-in, first-out (FIFO) buffer. The generated control signal sequence is not sent immediately but is first stored in this buffer for a precisely calculated buffer time. This buffer time equals the feedforward control lead time minus the total system delay. Only after this buffer time is the control signal sent to the actuator for execution. This eliminates timing jitter during the calculation process and ensures the accuracy of the output timing.

[0067] The closest prior art to this invention is a feedforward control system based on fixed delay compensation. This system measures or estimates a fixed system delay value and then simply advances the control signal by this fixed amount in the control timing. Compared to this prior art, the essential differences and innovations of this invention are: First, the delay management strategy is fundamentally improved. Existing technologies use static, fixed delay values ​​for compensation, which cannot adapt to the dynamic changes in system delay during operation. This invention, through periodic measurements and dynamic adjustments, can obtain the latest system delay time in real time and dynamically calculate the required feedforward control advance based on this changing value, achieving dynamic and adaptive delay management. Second, the timing accuracy control mechanism is more complete. Existing technologies lack the ability to handle timing jitter and can only perform simple fixed advance compensation. This invention, by introducing a buffer management mechanism, not only compensates for the average delay but also eliminates the timing jitter generated during calculation, ensuring the accuracy and stability of the output timing, which is impossible with simple fixed delay compensation. Third, the system safety design is more comprehensive. Existing technologies directly use the measured delay value for compensation without considering measurement errors and unexpected situations. This invention introduces a safety factor to ensure that the calculated feedforward control lead time always has a margin, absolutely meeting the core timing constraints, thus improving the reliability and safety of the system.

[0068] In summary, this invention solves the technical problem that fixed delay compensation methods cannot cope with system delay changes and timing jitter by adopting a comprehensive timing management strategy of dynamic calibration, adaptive adjustment and buffer management, thus providing more accurate and reliable timing assurance for forward-looking control.

[0069] In a preferred embodiment, in the intelligent vibration reduction and noise control method for rail transit, the adaptive filtering algorithm is executed in the generation step as follows: Block processing: The target reference signal sequence is divided into continuous data blocks, and each data block is processed uniformly; Frequency domain transformation: The reference signal sequence of the current data block is transformed to the frequency domain via a fast Fourier transform; Frequency domain filtering: In the frequency domain, the transformed reference signal is multiplied by the current frequency domain filter coefficients to directly generate a frequency domain control signal block; Time domain transformation: The frequency domain control signal block is transformed back to the time domain via an inverse fast Fourier transform to obtain the corresponding time domain control signal sequence; Block update: After obtaining the actual error signal of a complete data block, the filter coefficients are updated in one block in the frequency domain.

[0070] In the generation step, the adaptive filtering algorithm first divides the target reference signal sequence into consecutive data blocks, each containing a certain number of sampling points. The algorithm processes data blocks in batches, rather than processing them point-by-point as in traditional methods.

[0071] For each data block, the algorithm uses a Fast Fourier Transform to transform the time-domain reference signal to the frequency domain. In the frequency domain, the transformed reference signal is multiplied by the current frequency-domain filter coefficients using simple complex multiplication operations to directly generate the frequency-domain control signal block.

[0072] Then, the frequency domain control signal block is converted back to the time domain using an inverse fast Fourier transform to obtain the corresponding time domain control signal sequence. After obtaining the actual error signal of a complete data block, the algorithm performs a one-time block update of the filter coefficients in the frequency domain, rather than updating them at each sampling point.

[0073] The closest existing technology to this invention is an adaptive filtering algorithm based on time-domain least mean square. This technology uses a point-by-point sampling approach, requiring filtering calculations and coefficient updates at each sampling point. Compared to this existing technology, the essential differences and innovations of this invention are: First, a fundamental change in the processing method. Existing technologies use serial point-by-point processing, resulting in low computational efficiency and difficulty in meeting the demands of real-time processing of long sequences. This invention uses a block processing approach, packaging multiple sampling points into blocks for batch processing, significantly improving data processing throughput. Second, efficiency improvement through computational domain transformation. Existing technologies perform complex convolution operations in the time domain, resulting in high computational complexity. This invention, through domain transformation, converts time-domain convolution into frequency-domain complex multiplication, leveraging the efficiency of the Fast Fourier Transform to significantly reduce computational complexity. Third, a more efficient update strategy. Existing technologies update filter coefficients at each sampling point, resulting in frequent updates and high computational costs. This invention uses a block update strategy, updating coefficients only once every multiple sampling points, ensuring the algorithm's adaptability while significantly reducing the computational load. In summary, this invention effectively solves the problem of low computational efficiency in traditional adaptive filtering algorithms through block processing, frequency domain calculation, and block update strategies. It can meet real-time requirements while ensuring control accuracy, and provides an efficient computational scheme for generating high-quality control signals.

[0074] In a preferred embodiment, in the intelligent vibration reduction and noise reduction control method for rail transit, before sending the control signal sequence to the actuator in the output step, a feedforward compensation and protection step is added: Inverse model feedforward compensation: The control signal sequence is input into an actuator inverse model compensator; the actuator inverse model is a mathematical inverse model that is pre-established by the system to describe the dynamic relationship between the ideal input and the actual output of the actuator; the compensator outputs a pre-distorted compensated control signal, which, after driving the actuator, can make the actual output force or sound wave of the actuator approximate the requirements of the original control signal; Output amplitude limiting protection: The amplitude of the compensated control signal is monitored in real time. If its amplitude exceeds the saturation threshold of the linear operating region of the actuator, it is limited within the saturation threshold to prevent the actuator from entering the nonlinear saturation region and to protect the equipment safety.

[0075] Before the control signal is sent to the actuator, the system adds a feedforward compensation and protection step. First, the control signal sequence is input into an actuator inverse model compensator. This inverse model is pre-established using a system identification method and can accurately describe the dynamic relationship between the actuator's ideal input and actual output. The compensator pre-distorts the control signal and outputs a corrected, compensated control signal. The actuator inverse model is pre-established using a system identification method. Specifically, during system deployment or maintenance, a set of white noise or swept-frequency test signals covering its operating frequency band and amplitude range is input to the actuator, and its actual output (such as output force, acceleration, or sound pressure) is simultaneously acquired. Based on this input-output data pair, algorithms such as recursive least squares (RLS) or neural networks (such as time-delay neural networks, TDNN) are used to fit a mathematical inverse model that can accurately describe the dynamic relationship between the actuator's ideal input and actual output. This compensator uses this inverse model to pre-distort the control signal, outputting a compensated control signal. This signal drives the actuator, making the actuator's actual output approximate the requirements of the original ideal control signal. To ensure model accuracy, the system has a built-in model update mechanism that can periodically (e.g., monthly) or automatically re-identify and update the inverse model parameters when actuator performance drift is detected.

[0076] Simultaneously, the system monitors the amplitude of the compensated control signal in real time. When the signal amplitude exceeds the saturation threshold of the actuator's linear operating range, the system automatically limits the signal amplitude within this threshold. This amplitude limiting protection mechanism ensures that the actuator always operates within the linear range, avoiding nonlinear saturation. Specifically, the amplitude of the compensated control signal is monitored in real time, and a soft amplitude limiting algorithm (such as the hyperbolic tangent function tanh) is used to smooth and clip signals exceeding the actuator's linear operating range saturation threshold to avoid high-frequency distortion introduced by hard amplitude limiting. Simultaneously, the system monitors the actual output value of the actuator feedback loop (such as a built-in force sensor, current sensor, or laser displacement sensor) in real time and compares it with the command signal. If the instantaneous deviation exceeds 150% of the maximum allowable value, hard protection is immediately triggered, cutting off the output within 1ms. If the deviation continues (e.g., exceeding 50ms) beyond the tolerance (e.g., 20%), it is determined that the actuator may be malfunctioning or saturated, and the system will automatically trigger a degradation mode (e.g., gradually reducing the control gain until switching to passive monitoring mode) and report a maintenance alarm. The system continuously tracks the actuator's cumulative operating time, number of overruns, and thermal model temperature to perform health assessments and predictive maintenance alerts.

[0077] The control signal, after compensation and limiting, is finally sent to the actuator for execution, thus ensuring that the actual output of the actuator can accurately track the ideal control command of the algorithm.

[0078] The closest existing technology to this invention is an actuator control method based on the linear assumption. This technology assumes that the actuator has ideal linear characteristics and directly sends the control signal generated by the algorithm to the actuator for execution. Compared with this existing technology, the essential difference and innovation of this invention are as follows: First, it represents a fundamental breakthrough in control concept. Existing technologies ignore the nonlinear characteristics of the actuator, simply treating the actuator as an ideal linear device, leading to a significant deviation between the actual control effect and the theoretical expectation. This invention addresses the nonlinearity problem of the actuator directly, actively offsetting the nonlinear characteristics of the actuator through inverse model feedforward compensation, realizing a shift from avoiding the problem to actively solving it. Second, the compensation method is more precise and effective. Existing technologies lack compensation mechanisms for actuator characteristics or use simple linear compensation methods. This invention establishes an accurate actuator inverse model through system identification, enabling accurate pre-distortion processing of the control signal, fundamentally solving the control distortion problem caused by actuator nonlinearity. Third, the system protection mechanism is more comprehensive. Existing technologies often lack protection measures for the actuator's operating state, easily leading to the actuator entering the saturation nonlinear region or even being damaged. This invention ensures that the actuator always operates within the linear range through real-time monitoring and amplitude limiting protection, thus protecting equipment safety and guaranteeing the stability of control performance.

[0079] In summary, this invention effectively solves the problem of the influence of actuator nonlinearity on control accuracy through a dual mechanism of inverse model feedforward compensation and output limiting protection, achieving precise matching between the ideal algorithm command and the actual actuator output, and significantly improving the performance and reliability of the entire control system.

[0080] In a preferred embodiment, in the intelligent vibration reduction and noise control method for rail transit, the cross-domain modal correlation module in the correlation and fusion step is executed in the following ways: online saliency assessment: for each vibration channel, the coherence function values ​​of its signal and the target noise signal in the time and frequency domains are calculated in real time as the saliency index of the channel at the current moment; dynamic channel competition: all channels are sorted according to their saliency indices, and the top N channels with the highest indices are selected to form the current optimal vibration channel subset, where N is a preset fixed value; attention weight calculation: using an attention mechanism, the contribution weight of the channels in the current optimal vibration channel subset with the noise signal features is calculated only; weighted fusion: based on the calculated contribution weights, the signal features of the current optimal vibration channel subset are weighted and fused to generate the fused cross-domain feature vector.

[0081] It is important to note that the specific implementation of the cross-domain modal correlation module can adopt one of the following two preferred implementation methods, depending on the needs of the actual application scenario and the availability of prior knowledge. This method is suitable for scenarios with complex and variable environments or lacking complete prior knowledge. Its advantage is strong adaptability, enabling it to dynamically track changes in the vibration-acoustic coupling relationship.

[0082] In the correlation and fusion step, the system first performs an online saliency assessment. For each vibration channel, the system calculates the coherence function values ​​of its signal and the target noise signal in the time and frequency domains in real time. This coherence function value serves as the saliency index of that channel at the current moment, accurately reflecting the real-time correlation strength between vibration and noise.

[0083] Next, dynamic channel competition is conducted. The system sorts all channels according to their significance index and selects the top N channels with the highest index to form the current optimal subset of vibration channels. This N value is a preset fixed value, determined based on the system's computing power, to ensure that the computational load of subsequent processing is within a controllable range.

[0084] Then, an attention mechanism is employed to calculate the contribution weight of the noise signal features only for channels in the selected optimal vibration channel subset. Finally, based on the calculated contribution weights, the signal features of these selected channels are weighted and fused to generate a high-quality fused cross-domain feature vector.

[0085] The two cross-domain modal association implementation methods described above (based on prior pre-screening and based on online saliency dynamic competition) can be selected and applied through one of the following strategies: Strategy 1 (Parallel Redundant Operation): The system runs two sets of association modules simultaneously. The final output uses weighted fusion or selects the output with higher confidence, but this strategy consumes a lot of computational resources. Strategy 2 (Operating Condition Judgment Switching): The system presets a switching logic. When the system is just started, or the train is running on a fixed line section (such as a platform or parking lot), and historical data is sufficient, implementation method A (pre-screening) is used by default to pursue higher computational efficiency and stability. When the system detects a significant change in operating conditions (such as a large change in train speed or a track temperature change exceeding ±10℃) or the online evaluation shows that the average saliency index of the prior channel set is continuously lower than the threshold, it automatically switches to implementation method B (dynamic competition) to ensure the system's environmental adaptability. When switching between the two modes, the output of the fused feature vector needs to be smoothly transitioned to avoid abrupt changes in control commands. Strategy 3 (Manual Configuration): During system initialization, the engineer manually selects a mode based on the characteristics of the line and fixes it. This embodiment recommends using Strategy 2 (operating condition judgment and switching) to achieve a balance between intelligence and efficiency.

[0086] The closest existing technology to this invention is a method for pre-screening vibration channels based on fixed prior knowledge. This technology relies on offline-acquired static prior knowledge to pre-determine which vibration channels have a high correlation with noise, and then uses these channels consistently during system operation. Compared to this existing technology, the essential differences and innovations of this invention are: First, the screening mechanism is fundamentally transformed. Existing technologies use static and unchanging channel screening rules, which cannot adapt to environmental changes and system dynamic characteristics. This invention employs an online real-time calculation dynamic screening mechanism, accurately evaluating the current salience of each channel through coherence function values, and can automatically adapt to changes in various operating conditions. Second, the evaluation indicators are more scientific and accurate. Existing technologies rely on offline measured prior knowledge, which may become invalid over time. This invention uses the coherence function, a classic signal processing indicator, to evaluate the correlation strength between vibration and noise in real time and quantitatively, ensuring the accuracy and reliability of the screening results. Third, the system adaptability is significantly enhanced. Existing technologies suffer from screening failure when environmental conditions change. This invention, through an online dynamic competition mechanism, ensures that the system always focuses on the most important vibration channels, greatly improving adaptability and robustness under different operating conditions.

[0087] In summary, this invention solves the problem that static prior knowledge cannot adapt to dynamic environments by using online saliency assessment and a dynamic channel competition mechanism, thereby achieving intelligent and adaptive vibration channel screening and providing more accurate and reliable fusion features for subsequent processing.

[0088] In a preferred embodiment, the intelligent vibration reduction and noise reduction control method for rail transit further includes a dynamic prediction mode selection and compensation step in the prediction step. The dynamic prediction mode selection and compensation step includes: confidence assessment: the long short-term memory network model outputs a final value amplitude of... end Phase ƒ of final value end Simultaneously, it outputs the confidence score C of its predicted value; pattern fusion: a mixing factor α is set, the value of which is dynamically adjusted by the confidence score C; the decoder generates the final predicted sequence P. final The formula is: P final =α*P anchor +(1-α)*P recursive , where P anchor For a smooth transition sequence anchored at the final value, P recursive The sequence is generated by the model's recursive prediction; weighted compensation: when performing online error compensation, the feedback correction amount is weighted, and future time points that are similar to the current error frequency characteristics are given higher correction weights.

[0089] In the prediction step, the Long Short-Term Memory (LSTM) network model outputs the final value amplitude and final value phase, along with a confidence score for these predictions. This score reflects the model's assessment of the reliability of the current prediction results. Specifically, the LSM network model outputs the final value amplitude and final value phase... end Simultaneously, its internal Uncertainty Quantification module outputs a confidence score C (range 0-1) for its predicted values. This score is calculated based on the following three factors: ① Model inherent uncertainty: Calculated by using Monte Carlo Dropout technology for multiple forward propagations during the inference phase, based on the variance of the prediction results. The larger the variance, the lower the confidence score C. ② Input feature novelty: Calculating the Mahalanobis distance between the current input's cross-domain feature vector and the feature distribution of the training dataset, or using the autoencoder reconstruction error to determine whether the current input belongs to a distribution familiar to the model. The greater the distance or the larger the error, the lower C. ③ Recent prediction error: Statistically calculating the mean absolute error (MAE) between the model's predicted values ​​and the actual measured values ​​over a recent period (e.g., 50 time steps). The larger the error, the lower C.

[0090] The final confidence score C is determined by the weighted sum of the three components mentioned above and normalized to the interval [0, 1].

[0091] The system sets a blending factor, the value of which is dynamically adjusted by the confidence score. When the decoder generates the final predicted sequence, it no longer relies solely on a certain prediction pattern, but instead performs a weighted fusion of the smooth transition sequence generated by endpoint anchoring and the sequence generated by the model's recursive prediction. The weighting ratio is determined by the blending factor; when the confidence score is high, it favors the endpoint anchoring pattern, and when the confidence score is low, it favors the recursive prediction pattern.

[0092] During online error compensation, the system intelligently weights the feedback correction amounts. By analyzing the frequency characteristics of the current error, it assigns higher correction weights to future time points with similar characteristics, achieving precise and targeted compensation.

[0093] During weighted compensation, the feedback correction amount is weighted when performing online error compensation. Specifically, this is achieved by performing a short-time Fourier transform (STFT) on the prediction error e(t) at the current moment and extracting its dominant frequency component f. current And its amplitude. Similarly, STFT analysis is performed on the predicted sequence within the future Ta time period to obtain its frequency distribution at different time points. The dominant frequency and f at each future time point are calculated. currentThe cosine similarity of the frequencies or the absolute value of the frequency difference is used. A weight matrix is ​​calculated based on this similarity; the higher the similarity, the greater the correction weight assigned to that future time point. For example, the weight w(t) = exp(-λ*|f future (t)-f current |), where λ is the attenuation coefficient. This mechanism ensures that the correction energy is precisely directed to the future moments most likely to experience similar errors, thus improving compensation efficiency.

[0094] The closest existing technology to this invention is a single-mode prediction method. This technology either relies entirely on endpoint anchoring prediction or entirely on recursive prediction, lacking an adaptive switching mechanism between modes. Compared with this existing technology, the essential difference and innovation of this invention are: First, the prediction mode is innovatively broken. Existing technologies adhere to a single prediction mode, which either faces the problem of error accumulation or bears the risk of endpoint prediction. This invention creatively proposes multi-mode fusion prediction, dynamically adjusting the prediction strategy based on confidence scores, combining the advantages of both modes while avoiding their disadvantages. Second, the compensation mechanism is more intelligent. Existing technologies use uniform error compensation, treating all future time points equally. This invention uses frequency feature analysis for weighted compensation, which can achieve precise and targeted error correction, greatly improving the compensation efficiency and effect. Third, the system's fault tolerance is significantly improved. Existing technologies cannot self-adjust when prediction reliability declines. This invention, through confidence assessment and dynamic fusion, enables the system to have self-diagnosis and adaptive capabilities, and can intelligently adjust the prediction strategy when prediction reliability changes. In summary, this invention addresses the limitations of a single prediction mode through a triple mechanism of confidence assessment, multi-mode fusion, and intelligent weighted compensation, achieving a dual improvement in prediction accuracy and robustness, and providing more reliable prediction assurance for feedforward control.

[0095] In a preferred embodiment, in the intelligent vibration reduction and noise reduction control method for rail transit, the system delay calibration sub-step in the timing management step is performed in a non-disruptive manner as follows: Non-disruptive delay estimation: continuously monitor the control signal sequence and the actual effect signal fed back by sensors near the actuator; by calculating the cross-correlation function between the control signal and the actual effect signal, find the time offset corresponding to its maximum peak value, and take this time offset as the current total system delay time Td. current The estimated value; feedforward prediction: establish a regression model of system load and computing latency, predict the trend of computing latency changes based on the current core utilization and memory usage of the processor, and calculate the Td value. current The estimated value is fine-tuned by feedforward; feedback calibration: the feedforward-tuned Td is then adjusted. currentThe buffer management sub-step is run, and the deviation between the actual control effect and the expected target is continuously compared; based on the deviation, a PID controller is used to calibrate the parameters of the regression model online.

[0096] In the timing management step, system delay calibration is performed in a non-disruptive manner. The system continuously monitors the control signal sequence and the actual effect signals fed back by sensors near the actuator. By calculating the cross-correlation function between these two signals, the time offset corresponding to the maximum peak is found, and this time offset is used as an estimate of the current total system delay time. During the non-disruptive delay estimation process, during normal system operation, the control signal sequence and the acceleration / error microphone signals near the actuator are continuously buffered in a loop. Every 100 milliseconds is a calculation cycle. After applying a Hanning window to the two signals within this cycle, the cross-correlation function is calculated to find the time offset corresponding to the maximum peak. A median filtering algorithm is then used to smooth the estimation results of five consecutive cycles, finally outputting a stable and disturbance-resistant Td. current Estimated value.

[0097] The system also establishes a load-latency regression model, predicting the trend of latency changes based on the current core utilization and memory usage of the processor, and fine-tuning the latency estimate using feedforward. Finally, the regression model parameters are calibrated online using a PID controller, and the model accuracy is continuously optimized based on the deviation between the actual control effect and the expected target.

[0098] The closest prior art to this invention is a delay calibration method based on test signal injection. This technique requires periodically injecting dedicated test signals during system operation and calibrating the system delay by measuring the transmission delay of the test signals. Compared with this prior art, the essential differences and innovations of this invention are: First, the calibration method is fundamentally improved. Existing techniques require the injection of external test signals, which interferes with the normal control process. This invention uses non-disruptive delay estimation, utilizing signals from the normal control process for correlation analysis, completely avoiding interference from test signals to system operation. Second, the delay prediction capability is significantly improved. Existing techniques can only measure delay and cannot predict delay change trends. This invention, by establishing a regression model between load and delay, can predict delay changes based on system resource usage, achieving proactive adjustments. Third, the system adaptability is stronger. The calibration results of existing techniques are prone to drift over time. This invention, through an online calibration mechanism, can continuously maintain the accuracy of delay estimation and adapt to performance changes during long-term system operation.

[0099] In summary, this invention solves the problems of test signal interference and dynamic delay changes through an innovative combination of non-disruptive delay estimation, feedforward prediction, and online calibration, achieving accurate, non-disruptive, and adaptive management of system delay, and providing a more reliable technical guarantee for timing control.

[0100] In a preferred embodiment, in the intelligent vibration reduction and noise reduction control method for rail transit, the adaptive filtering algorithm in the generation step is executed in the following optimized manner: Hybrid domain processing: The target reference signal sequence is divided into continuous data blocks; for each data block, it is further divided into a first half and a second half; the frequency domain block processing method is used to generate control signals for the second half of the data blocks; for the first half of the data blocks, a time domain filtering algorithm is used for parallel computation to significantly reduce the processing delay caused by waiting for the entire data block; Adaptive block length adjustment: The main frequency of the target reference signal is monitored in real time; the optimal data block length is dynamically selected based on the Nyquist sampling theorem and the available computing resources of the current system; when the main frequency is high, a smaller block length is automatically selected to reduce the delay; when the main frequency is low, a larger block length is automatically selected to improve computational efficiency; Error compensation filtering: After the time domain transformation, a digital post-filter is applied to the generated time domain control signal sequence. The frequency response characteristics of this filter are designed to be opposite to the distortion characteristics introduced by circular convolution during the frequency domain block processing, in order to compensate for the aforementioned linear convolution error.

[0101] In the generation step, the adaptive filtering algorithm employs a hybrid domain processing strategy. First, the target reference signal sequence is divided into continuous data blocks, each of which is further divided into a first half and a second half. The second half of the data block is processed using frequency domain block processing to generate the control signal, while the first half is processed in parallel using a time domain filtering algorithm, significantly reducing the processing latency of waiting for the entire data block to be generated.

[0102] The system monitors the dominant frequency of the target reference signal in real time and dynamically selects the optimal data block length based on the Nyquist sampling theorem and available computing resources. When the dominant frequency is high, a smaller block length is automatically selected to reduce latency; when the dominant frequency is low, a larger block length is automatically selected to improve computational efficiency.

[0103] After time-domain transformation, a digital post-filter is applied to the generated time-domain control signal sequence. The frequency response characteristics of this filter are specifically designed to compensate for the distortion introduced by circular convolution during frequency domain block processing, effectively eliminating linear convolution errors.

[0104] The closest existing technology to this invention is a pure frequency domain block processing method. This technique processes the entire data block in the frequency domain, which improves computational efficiency but introduces significant block latency and linear convolution errors.

[0105] Compared with existing technologies, the essential differences and innovations of this invention are as follows: First, the processing architecture is significantly innovative. Existing technologies employ a single frequency domain processing mode, which cannot balance efficiency and real-time performance. This invention creatively proposes a hybrid domain processing architecture, combining the speed advantage of time domain processing with the efficiency advantage of frequency domain processing, achieving an optimal balance between latency and efficiency. Second, the block length selection is intelligent. Existing technologies use fixed block lengths, which cannot adapt to changes in signal characteristics. This invention dynamically adjusts the block length by monitoring signal frequency characteristics in real time, ensuring the system always operates under optimal parameters. Third, the error compensation mechanism is improved. Existing technologies neglect the error problem caused by circular convolution. This invention effectively compensates for the distortion introduced by frequency domain processing through a specially designed post-filter, ensuring the accuracy of the control signal. In summary, this invention, through hybrid domain processing, dynamic block length adjustment, and error compensation, effectively solves the problems of large latency and large error in pure frequency domain block processing methods, significantly improving real-time performance and accuracy while maintaining computational efficiency, providing a better solution for control signal generation.

[0106] To ensure long-term stable operation in complex rail transit environments, this invention integrates multiple reliability assurance mechanisms at the architecture level.

[0107] (1) Redundancy and fault tolerance of the sensing layer: The sensor array adopts a distributed deployment and has built-in self-diagnostic functions (such as signal amplitude range check and DC bias detection). When an individual sensor fails or data is abnormal, the system can automatically remove it from the fusion calculation and rely on the remaining sensors to reconstruct the signal through Kalman filtering or interpolation algorithms to ensure the continuity and robustness of the sensing system.

[0108] (2) Confidence management and degradation strategy of prediction layer: The prediction module outputs the real-time confidence score C of its prediction results. When C is lower than the preset threshold (e.g., 0.6), it indicates that the prediction is unreliable due to strong interference or model mismatch. The system can automatically and smoothly switch to a more conservative mitigation control strategy (e.g., only control the low-frequency band that has been verified to be effective, or reduce the control gain) to prevent control instability or positive feedback due to incorrect prediction.

[0109] (3) Execution layer closed-loop protection and health monitoring: Real-time sensor feedback of actuator output force / displacement / current is introduced and compared with the command signal to form a fast local closed loop. Once an output abnormality is detected (such as saturation, distortion, or deviation exceeding the tolerance for 100ms), a graded safety response mechanism is immediately triggered (such as amplitude limiting, gain reduction, or even safety shutdown) to ensure the safety of the actuator itself and the controlled structure. At the same time, the system records the actuator's working status and performs health assessment and life prediction.

[0110] (4)Dynamic tolerance management at the timing level: The timing management module continuously monitors and compensates for the jitter of system latency. Through the feedforward-feedback composite control strategy, even if the computing time slightly increases due to processor load fluctuations, the dynamically adjusted advance Ta_required and buffer time T_buffer can always ensure that the core condition of Td < Ta holds, thus avoiding the positive feedback oscillation caused by timing disorder at the root.

[0111] Through these designs, the present invention constructs a highly reliable intelligent control system with the capabilities of fault self-diagnosis, performance self-evaluation, and strategy self-adjustment, ensuring long-term stable operation in a real industrial environment. Embodiment

[0112] This embodiment provides an intelligent shock and noise reduction control system for rail transit, which is deployed in a subway tunnel section. The system hardware includes: 64 acceleration vibration sensors and 32 noise sensors are evenly distributed within the range of a 100-meter-long track and tunnel wall; the central processing unit uses a multi-core processor with AI acceleration function; the actuator array consists of 16 electro-hydraulic actuators.

[0113] When the system works, first, the vibration and noise signals are synchronously collected through the sensor array. The cross-domain modal correlation module uses an online significance evaluation method to calculate the coherence function values of each vibration channel and the noise signal in real time, and selects the top 20 channels with the highest coherence values to form an optimal channel subset. The lightweight multi-head attention mechanism is used to calculate the contribution weights of these channels in parallel and generate a fused feature vector. Specifically, when the system works, first, the vibration and noise signals are synchronously collected through the sensor array. The data is transmitted to the central processing unit through the TSN network, and its fixed end-to-end transmission delay is less than 2 milliseconds. The cross-domain modal correlation module runs on the NPU of the processor, and the single processing time is less than 5 milliseconds. The long short-term memory network model has been pre-trained and quantified using historical track data, and the single forward prediction time is less than 10 milliseconds. The timing management module measures the total system latency Td current to be 150 ± 5 milliseconds (including all links of sensing, transmission, calculation, and actuator response), and accordingly sets the feedforward control advance Ta required to be 180 milliseconds, leaving sufficient margin for latency jitter. The total computing time of the entire control loop is strictly controlled within 50 milliseconds, thus stably meeting the core timing constraint of Td < Ta.

[0114] After receiving the fused features, the Long Short-Term Memory (LSTM) network model first predicts the amplitude and phase of the final vibration value 200 milliseconds in the future, with a confidence score of 0.85. The decoder uses this final value as a constraint to generate a smooth transition sequence, while simultaneously performing weighted compensation based on real-time prediction errors. The timing management module estimates the current total system delay to be 150 milliseconds using perturbation-free delay estimation, and accordingly sets the feedforward control lead to 180 milliseconds, along with a 30-millisecond buffer time.

[0115] Adaptive filtering employs hybrid-domain processing: the reference signal is divided into 128 data blocks, with the first 64 points using time-domain filtering and the last 64 points using frequency-domain processing. The block length is dynamically adjusted based on the signal's dominant frequency; a block length of 64 points is used when the dominant frequency exceeds 100Hz, and a block length of 256 points is used when it is below 50Hz. A post-filter compensates for frequency-domain processing distortion.

[0116] The actuator inverse model compensator pre-distorts the control signal, and the output limiting protection module restricts the signal amplitude within the actuator's linear operating range. Finally, the control signal, after buffering and delay, drives the actuator to generate a counteracting force.

[0117] Comparative Example 1: Traditional Passive Control A passive control system using high-damping fasteners and track bed vibration isolation pads is employed. This system is based on the principle of physical vibration isolation, without sensors or processors, relying on material properties to achieve wide-band vibration attenuation. However, its vibration reduction effect is fixed, unable to adapt to different vehicle speeds and load conditions, has poor suppression of low-frequency vibrations, and lacks noise reduction functionality.

[0118] Comparative Example 2: Basic Active Control System A feedback-based active control system is employed. The system uses eight vibration sensors, a traditional PID controller, and electromagnetic actuators. Upon detecting vibration, the system generates an anti-phase control signal. However, due to a 50-millisecond system delay, the control effect on vibrations above 100Hz is significantly reduced, and control instability frequently occurs due to phase lag.

[0119] Comparative Example 3: Control System Based on Fixed Channel Selection This system does not employ online saliency assessment. Instead, it uses a fixed channel selection strategy, determining 20 vibration channels through offline testing during the system design phase. These channels are selected based on acoustic transfer function analysis under initial operating conditions. During operation, the system consistently uses these 20 preset channels without adjustment based on environmental conditions. The attention network employs a standard-scale structure, performing weight calculations and feature fusion on the signals from all selected channels. This system lacks an online channel evaluation mechanism and cannot adjust the channel selection strategy based on real-time signal characteristics.

[0120] System latency performance test Total system delay time Tdcurrent It is the total time established for the stable build-up from the acceleration sensor signal input to the actuator output force. This index is tested under the condition that the central processing unit (CPU) load rate is 60% (simulating typical operating load).

[0121] Test method: Inject a Heaviside step signal with a sudden change in amplitude into the sensor data stream, use a high-speed oscilloscope (bandwidth > 100 MHz) to synchronously capture the input signal and the response signal of the actuator output force sensor, and measure the time difference between the two. The test is repeated 100 times.

[0122] Test result: The total system delay time Td is measured. current It is (150 ± 5) ms (mean ± standard deviation). This delay includes: the sensor signal transmission and synchronization delay is less than 2 ms; the data processing and algorithm calculation delay is (125 ± 5) ms, and its fluctuation mainly stems from the calculation cycle jitter of the adaptive filter block processing; the actuator (electro-hydraulic) response delay is (23 ± 1) ms. The load impact test shows that when the CPU load rate varies between 20% and 90%, there will be a ±10 ms fluctuation in the calculation delay part, but through the dynamic adjustment of the timing management module, the core constraint of Td < Ta can always be satisfied eventually.

[0123] Performance test conditions and results Under the standard test conditions (train speed 80 km / h, axle load 16 tons, seamless track in straight section, measurement point in tunnel), the vibration and noise reduction performance of the system of the embodiment of the present invention is tested according to the following method.

[0124] Measurement point layout: The vibration measurement point is located 1.5 meters away from the track center line on the surface of the ballast bed, and a vibration sensor certified by ISO 8041-1 is used; the noise measurement point is located on the side of the tunnel wall, 1.2 meters above the rail surface and 3.5 meters away from the track center line, and an IEC61672 Class1 sound level meter is used.

[0125] Measurement method: Collect data within 10 seconds before and after the train passes, and the insertion loss is used as the evaluation index. That is: vibration / noise reduction effect = 20 * log10(pre-control vibration acceleration / noise pressure / post-control vibration acceleration / noise pressure). The test is repeated 10 times and the average value is taken.

[0126] Test result: Vibration reduction effect: In the main vibration frequency band of 20 - 200 Hz, the insertion loss of the Z vibration level reaches (28 ± 2) dB (mean ± standard deviation).

[0127] Noise reduction effect: In the 1 / 3 octave frequency range of 63 - 4000 Hz, the insertion loss of the A-weighted sound pressure level reaches (18 ± 1.5) dB.

[0128] Comparative Example 1 uses high-damping fasteners and track bed vibration isolation pads, achieving an overall vibration reduction of only 15dB and a noise reduction of less than 5dB. It is primarily effective for mid-to-high frequencies (approximately 12dB vibration reduction above 500Hz), offering almost no suppression of low-frequency vibrations (below 50Hz). The Example 1 automatically adapts to different vehicle speeds (60-100km / h) and load conditions through an intelligent algorithm, with performance fluctuations less than 2dB. The performance of Comparative Example 1 depends entirely on fixed physical parameters. When the vehicle speed increases from 60km / h to 100km / h, the vibration reduction effect decreases by approximately 40% due to the change in excitation frequency, and it cannot cope with changes in track structure characteristics caused by aging. The Example 1 possesses complete vibration and noise coordination control capabilities, simultaneously handling structural vibration and airborne noise. Comparative Example 1 can only suppress vibrations propagating through the structure, offering no direct noise reduction effect on airborne noise, requiring the installation of additional sound barriers.

[0129] In dynamic performance testing, the embodiment employs look-ahead predictive control, achieving a 25dB suppression effect on low-frequency vibrations below 100Hz and a stable 15-20dB control effect on noise above 500Hz. Comparative Example 2 uses feedback control, but due to system delay (50ms) and signal processing limitations, it can only effectively control vibrations in the 50-200Hz frequency band (approximately 12dB), with a control effect of less than 5dB on high-frequency components and significant phase lag. The embodiment ensures system stability through timing management and buffer control, exhibiting no instability under various operating conditions. Comparative Example 2's PID control frequently exhibits overshoot oscillations when handling complex vibration signals, easily causing system instability near the resonant frequency, requiring manual intervention to adjust parameters. The embodiment's control bandwidth covers 20-2000Hz, with a response time of less than 100ms. Comparative Example 2's effective control bandwidth is only 50-500Hz, with a response time exceeding 150ms, and a significant delay in response to sudden impact vibrations, resulting in a control performance decrease of approximately 50%.

[0130] Under the same test conditions, the system in the embodiment showed a significant advantage in environmental adaptability compared to Comparative Example 3. When the track temperature rose from 20°C to 35°C, the vibration reduction effect of Comparative Example 3 decreased by approximately 40% due to its fixed channel selection strategy. In contrast, the system in the embodiment automatically adjusted the selection of important channels through online significance assessment, thus maintaining stable performance.

[0131] In the comprehensive performance test, the vibration reduction effect of the system in the example was improved by about 50% compared with the control example 3, and it maintained excellent and stable control performance under various operating conditions.

[0132] The system successfully achieved forward-looking and precise control. To verify its reliability, targeted tests were conducted: Sensor fault simulation test: Randomly shield 20% of the vibration sensor signals. The system performance degrades by less than 3 dB and there is no instability phenomenon, indicating that it has good sensing fault tolerance ability.

[0133] Anti-interference test: Apply additional impact vibrations (simulating construction interference) that are not caused by train operation beside the track. The system can effectively filter this interference through the above error threshold mechanism, and the prediction sequence is not significantly affected, and the control remains stable.

[0134] Long-term operation test: The system runs continuously for 720 hours. The system delay Td_current recorded by the timing management module fluctuates between 145 ms and 155 ms, and the dynamically adjusted lead Ta_required is always stable at 180 ms. The core timing constraint Td < Ta holds 100% during the entire test period, and no control failure due to timing errors occurs.

[0135] Actuator protection test: Manually set the actuator output to saturation. The system detects an abnormal deviation within 50 milliseconds and triggers the degradation mode, effectively protecting the actuator equipment.

[0136] Although the embodiments of the present invention have been disclosed as above, it is not limited to the applications listed in the specification and embodiments. It can be fully applied to various fields suitable for the present invention. For those familiar with the field, additional modifications can be easily achieved. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to specific details.

Claims

1. A rail transit intelligent shock absorption and noise reduction control method, characterized in that, The method performs the following steps: A collection step: synchronously acquiring vibration signals collected by a plurality of acceleration sensors arranged along the track and noise signals collected by a microphone; An association and fusion step: inputting the vibration signals and noise signals into a cross-domain modal association module; the cross-domain modal association module is based on an attention mechanism, calculates the contribution weight of each vibration signal channel to the noise signal feature, and performs weighted fusion on the vibration signal feature according to the contribution weight, and outputs a fused cross-domain feature vector; A prediction step: inputting the cross-domain feature vector into a pre-trained long short-term memory network model, which outputs a prediction sequence of the target vibration main frequency, amplitude and phase within a feedforward control advance time Ta, and a prediction sequence of the target point noise sound pressure level; A synthesis step: synthesizing a target reference signal sequence corresponding to the feedforward control advance time Ta according to the prediction sequence; A generation step: inputting the target reference signal sequence into an adaptive filtering algorithm to generate a control signal sequence with equal amplitude and opposite phase; An output step: outputting the control signal sequence to the actuator in real time, wherein the total system delay time Td from the start of calculation to the action of the actuator is strictly less than the feedforward control advance time Ta, so as to realize prospective cancellation of the target vibration and noise.

2. The intelligent vibration and noise reduction control method for rail transit according to claim 1, wherein In the association and fusion step, the cross-domain modal association module is executed in the following manner: A pre-screening sub-step: based on prior knowledge of the track structure, a specific vibration channel subset with an acoustic contribution correlation degree higher than a preset threshold is selected from all vibration signal channels; A parallel calculation sub-step: a lightweight multi-head attention mechanism is used to calculate the contribution weight between the signal in each channel in the specific vibration channel subset and the noise signal feature in parallel; A weighted fusion sub-step: the signal features of the specific vibration channel subset are weighted and fused according to the calculated contribution weight to generate the fused cross-domain feature vector.

3. The rail transit intelligent shock absorption and noise reduction control method of claim 1 or 2, characterized in that, In the prediction step, the long short-term memory network model predicts in the following manner: First stage rough prediction: according to the input cross-domain feature vector, the final value amplitude A of the vibration main frequency at the end of the feedforward control lead time Ta is predicted end and the final value phase f end ; Second stage sequence generation: with the end value amplitude A end and end value phase f end as strong constraints, a smooth transition sequence of amplitude and phase from the current time to time Ta is generated by an internal decoder that meets the end value constraints; Online compensation correction: the error between the predicted value at the current time and the actual sensing measurement value at the time is obtained in real time, and the error is used as a feedback correction amount and injected into the generation process of the decoder to dynamically fine-tune the future prediction sequence that has not been output.

4. The rail transit intelligent shock absorption and noise reduction control method of claim 3, wherein, The method further includes a timing management step, which is run in parallel with the collection to output steps and includes: System delay calibration: inject a test signal to the control loop when the system is starting or idle, measure and record the total system delay time Td from the signal input to the actuator response output current ; Adaptive adjustment of advance: the measured Td current is multiplied by the preset safety factor to obtain the current required feedforward control advance time Ta required , and the prediction length in the prediction step is dynamically adjusted to be equal to Ta required ; Buffer management: a first-in-first-out buffer is set before the output step; the control signal sequence is temporarily stored in the buffer after being generated, and waits for a fixed buffer time T buffer Then, it is sent to the actuator; the buffer time T buffer According to the formula T buffer = Ta required -Td current is calculated to eliminate the jitter of calculation delay.

5. The intelligent vibration and noise reduction control method for rail transit according to claim 4, wherein In the generation step, the adaptive filtering algorithm is executed in the following manner: Block processing: the target reference signal sequence is divided into continuous data blocks, and each data block is uniformly processed; Frequency domain transformation: the reference signal sequence of the current data block is converted to the frequency domain through fast Fourier transform; Frequency domain filtering: in the frequency domain, the transformed reference signal is directly multiplied by the current frequency domain filter coefficients to generate the control signal block in the frequency domain; Time domain conversion: the frequency domain control signal block is converted back to the time domain through inverse fast Fourier transform to obtain the corresponding time domain control signal sequence; Block update: after obtaining the actual error signal of a complete data block, the filter coefficients are updated in the frequency domain.

6. The rail transit intelligent shock absorption and noise reduction control method of claim 5, wherein, In the output step, before sending the control signal sequence to the actuator, a feedforward compensation and protection step is added: Inverse model feedforward compensation: the control signal sequence is input into an actuator inverse model compensator; the actuator inverse model is a mathematical inverse model pre-established by system identification to describe the dynamic relationship between the ideal input and the actual output of the actuator; the compensator outputs a pre-distorted compensated control signal, which drives the actuator and makes the actual output force or sound wave of the actuator approach the requirements of the original control signal; Output amplitude limiting protection: the amplitude of the compensated control signal is monitored in real time, and if its amplitude exceeds the saturation threshold of the linear working area of the actuator, it is limited within the saturation threshold to prevent the actuator from entering the nonlinear saturation area and protect the safety of the equipment.

7. The rail transit intelligent shock absorption and noise reduction control method of claim 1, wherein, In the correlation and fusion step, the execution mode of the cross-domain modal correlation module includes: Online saliency evaluation: for each vibration channel, the coherence function value of its signal and the target noise signal in the time and frequency domains is calculated in real time as the saliency index of the channel at the current time; Dynamic channel competition: all channels are sorted according to their saliency indexes, and the top N channels with the highest indexes are selected to form the current optimal vibration channel subset, where N is a predetermined fixed value; Attention weight calculation: using the attention mechanism, only the channels in the current optimal vibration channel subset are calculated for their contribution weights with the noise signal features; Weighted fusion: according to the calculated contribution weights, the signal features of the current optimal vibration channel subset are weighted and fused to generate the fused cross-domain feature vector.

8. The rail transit intelligent shock absorption and noise reduction control method of claim 3, wherein, In the prediction step, a dynamic prediction mode selection and compensation step is also included, which includes: Confidence evaluation: the long short-term memory network model outputs a confidence score C of its predicted value at the same time as the terminal value amplitude A end and the terminal value phase f end ​ Mode fusion: a mixing factor a is set, whose value is dynamically adjusted by the confidence score C; the decoder generates a final prediction sequence P final The formula is: P final = a * P anchor + (1-a) * P recursive , where P anchor is a smooth transition sequence anchored at the final value, and P recursive is a sequence generated by model recursive prediction; Weighted compensation: when performing online error compensation, the feedback correction amount is weighted to give higher correction weights to future time points with similar error frequency characteristics.

9. The rail transit intelligent shock absorption and noise reduction control method of claim 4, wherein, The system delay calibration sub-step in the timing management step is executed in the following non-disturbance manner: Non-disturbing delay estimation: continuously monitor the control signal sequence and the actual effect signal fed back by the sensor near the actuator; find the time offset corresponding to the maximum peak of the cross-correlation function between the control signal and the actual effect signal, and take this time offset as the estimated value of the current system total delay time Td current ​ Feedforward prediction: a regression model of system load and computation delay is established, the trend of computation delay is predicted according to the current core utilization and memory occupancy of the processor, and the estimated value is fine-tuned by feedforward current ​ Feedback calibration: Td after fine-tuning of feedforward current The buffer management sub-step is run and the deviation of the actual control effect from the expected target is continuously compared; based on this deviation, a PID controller is used to calibrate the parameters of the regression model online.

10. The rail transit intelligent shock absorption and noise reduction control method of claim 5, wherein, The adaptive filtering algorithm in the generation step is executed in the following optimization manner: Mixed domain processing: the target reference signal sequence is divided into consecutive data blocks; for each data block, it is further divided into a first half and a second half; the second half data block is processed in the frequency domain to generate the control signal; the first half data block is processed in parallel using the time domain filtering algorithm to significantly reduce the processing delay caused by waiting for the entire data block; Block length self-adaptive adjustment: real-time monitoring of the dominant frequency of the target reference signal; according to the Nyquist sampling theorem and the current available computing resources of the system, dynamically selecting the optimal data block length; when the dominant frequency is high, a smaller block length is automatically selected to reduce the delay; when the dominant frequency is low, a larger block length is automatically selected to improve the computing efficiency; Error compensation filtering: after the time domain conversion, a digital post-filter is applied to the generated time domain control signal sequence, and the frequency response characteristics of the filter are designed to be opposite to the distortion characteristics introduced by the circular convolution in the frequency domain block processing process, so as to compensate for the linear convolution error.