Subway-oriented intelligent detection monitoring method and system

Abstract

The invention discloses a subway-oriented intelligent detection monitoring method and system, and belongs to the technical field of detection. The method comprises the following steps: acquiring a sensor signal; preprocessing the sensor signal to obtain a preprocessed sensor signal; performing feature extraction on the preprocessed sensor signal to obtain a feature value; comparing the characteristic value with a double-layer threshold value to obtain a diagnosis result; and compressing and transmitting the diagnosis result. According to the intelligent vibration noise sensor and the data processing method thereof, the intelligent vibration noise sensor and the data processing method thereof are used for complex environments such as subways, and the sensor realizes high-fidelity acquisition, edge end intelligent processing and low-bandwidth reliable transmission of vibration noise data through deep fusion of a multi-mode sensing architecture and an efficient data compression transmission technology.

Description

Technical Field

[0001] This invention relates to the field of detection technology, and specifically to an intelligent detection and monitoring method and system for subways. Background Technology

[0002] With the deepening of urbanization in my country, the subway, as the main artery of urban public transportation, has seen its operational scale and passenger volume continuously increase, placing unprecedentedly high standards and stringent requirements on operational safety, passenger comfort, and maintenance efficiency. The subway system is a complex dynamic system composed of tracks, tunnels, and vehicles, operating under high intensity and high load conditions for extended periods. Its structural health and operational status directly affect the travel safety of millions of passengers. Traditional maintenance models mainly rely on manual periodic inspections and offline spot checks, which are not only inefficient and costly but also inherently subjective and unable to detect sudden faults in real time, making them unsuitable for the development needs of modern subway networks and high-density operations.

[0003] Currently, the monitoring of track, tunnel structures, and train operation status in subway operation and maintenance generally suffers from the following pain points: First, monitoring methods are fragmented. Key parameters such as vibration, noise, and deformation are usually collected independently by single-function sensors deployed in different locations and of different models. The data is isolated and asynchronous, making it difficult to form comprehensive diagnostic conclusions and failing to fully reflect the overall health status of the system. Second, the real-time and intelligent nature of data processing is insufficient. Most of the massive monitoring data needs to be transmitted back to the backend for preliminary analysis, resulting in high latency and bandwidth pressure, making it impossible to achieve real-time edge intelligent judgment and early fault warning for critical events such as train passage. Third, existing sensor networks have complex cabling and power supply difficulties, especially in harsh environments such as tunnels. Laying cables and power lines for a large number of sensors is extremely costly and inconvenient to maintain, limiting the density and coverage of the monitoring network. Finally, the data chain is not smooth. The entire process from front-end acquisition to back-end display suffers from problems such as inconsistent data protocols, low transfer efficiency, chaotic storage, and unstable transmission, preventing decision-makers from obtaining key information in a timely, clear, and accurate manner, affecting operation and maintenance efficiency and train operation safety.

[0004] Against this backdrop, intelligent monitoring technology has become an inevitable trend for ensuring subway safety and achieving predictive maintenance. Currently, although some subway lines have attempted to introduce online monitoring methods, three major pain points are prevalent: First, low system integration. Monitoring of different parameters such as vibration, noise, and deformation usually uses independent subsystems, resulting in redundant sensor deployments, inconsistent data protocols, and the formation of "information silos," making it difficult to perform multi-source information fusion and comprehensive diagnosis. Second, lagging data processing capabilities. Most of the massive monitoring data needs to be transmitted to a central server for processing, which puts enormous pressure on network bandwidth and causes significant delays, making it impossible to perform real-time edge intelligent analysis of critical events such as train passage, thus missing the golden window for fault early warning. Third, the challenges of power supply and cabling for the sensing layer. The complex tunnel environment makes laying power supply and communication lines for a large number of wired sensors a massive project with extremely difficult maintenance, severely restricting the coverage density and long-term stability of the monitoring network.

[0005] Therefore, developing a highly integrated, intelligent, self-consistent, and wirelessly interconnected intelligent detection and monitoring system is of vital importance for breaking through existing technological bottlenecks and promoting the transformation and upgrading of subway operation and maintenance models towards digitalization and intelligence. Summary of the Invention

[0006] The purpose of this invention is to provide an efficient intelligent detection and monitoring method and system for subways. By systematically integrating multiple intelligent sensors and multifunctional dedicated sensing nodes, an integrated sensing network is constructed to comprehensively perceive key state parameters such as vibration and noise. This system innovatively employs a signal relay module based on edge computing, pushing computing power down to the network edge to achieve localized real-time data processing and feature extraction, greatly reducing the burden of data transmission and storage. Finally, through hybrid wireless networking technology, the processed high-value information is seamlessly connected to the backend display system, providing operation and maintenance managers with real-time, clear, and comprehensive decision support. This fundamentally transforms maintenance from passive, experience-based methods to proactive, data-driven precision maintenance, laying a solid technical foundation for building a safe, efficient, and green smart subway.

[0007] To address the aforementioned technical problems, this invention provides an intelligent detection and monitoring method for subways, comprising the following steps: Acquire sensor signals; The sensor signal is preprocessed to obtain the preprocessed sensor signal; Feature extraction is performed on the preprocessed sensor signal to obtain feature values; The diagnostic results are obtained by comparing the feature values ​​with the two-layer threshold. The diagnostic results are compressed and transmitted.

[0008] Preferably, acquiring sensor signals specifically includes the following steps: The sensor signals are continuously monitored while the main controller is in deep sleep mode; When the sensor signal exceeds the dynamic threshold, the main controller is activated. After waking up the main controller, the sensor signals are analyzed, and a full start is initiated if the analysis result is abnormal. After full operation, regular inspections will be conducted.

[0009] Preferably, the formula for calculating the dynamic threshold is: in: For dynamic thresholds; This represents the baseline average value of the sensor signal; The noise standard deviation of the sensor signal; This is a configurable sensitivity coefficient; The formula for calculating the warning threshold is: in: This is the warning threshold; The inspection interval of the timed inspection The calculation formula is: in: Basic inspection interval; This is the maximum adjustment range coefficient; The equipment health index (derived from historical data analysis; the higher the value, the healthier the equipment). For health and safety thresholds; To adjust the width parameter of the interval; It is a hyperbolic tangent function used for smooth transitions.

[0010] Preferably, the sensor signal is preprocessed to obtain a preprocessed sensor signal, specifically including the following steps: An improved adaptive Kalman filter algorithm is used to preprocess the sensor signal: State prediction: Covariance prediction: Process noise covariance Adaptive adjustment as the signal gradient changes: in: The initial process noise covariance; Let k be the gradient of the signal amplitude at time k; It is a smoothing factor; When the signal changes drastically, reduce More trust in predicted values ​​to suppress transient interference; when the signal is stable, increase This allows for greater trust in observations and improved tracking accuracy.

[0011] Preferably, feature extraction is performed on the preprocessed sensor signal to obtain feature values, specifically including the following steps: A multi-scale deep feature fusion algorithm is used to extract features from the preprocessed sensor signal: Temporal augmentation features: including root mean square (RMS) and kurtosis. Waveform index SI, peak index and impact index ; , Frequency domain intelligent features: the spectrum after FFT transformation Perform calculations; Spectral centroid FC: reflects the location where spectral energy is concentrated; Spectral variance FV: reflects the degree of dispersion of the spectrum; After the central control module performs real-time preprocessing on the raw data, it executes an enhanced feature extraction algorithm. In addition to calculating the effective value of vibration velocity, peak acceleration, and equivalent sound level, it adds the following: Time-frequency domain hybrid features: Extracting wavelet packet energy entropy through wavelet packet transform Characterizing the non-stationary properties of a signal: , .

[0012] Preferably, the spectrum The calculation method is as follows: For the preprocessed time-domain signal x(n), n=0,1,...,L-1, where L is the number of sampling points; apply a window function w(n) to reduce spectral leakage: The Hanning window is chosen as the window function: Performing an FFT on the windowed signal xw(n) yields the complex spectrum X(k): Where: k is the spectrum line index, corresponding to the digital frequency. L is the number of FFT points; The amplitude information is extracted from the complex spectrum X(k) to obtain the amplitude spectrum A(k): in: , where is the modulus of the complex spectrum; Convert the spectral line index k to the actual physical frequency fk: Where: fs is the sampling frequency; fk is the actual frequency corresponding to the kth spectral line; Spectral sequence calculated from frequency domain features: Where N = L / 2, and Ak is the amplitude sequence used in the frequency domain feature calculation.

[0013] Preferably, the feature value is compared with a two-layer threshold to obtain the diagnostic result, specifically including the following steps: Compare the feature values ​​with the adaptive physical threshold; The adaptive physical threshold for: in: It is the threshold under standard conditions. and It is the reference frequency and temperature. and It is a correction factor; The frequency domain intelligent feature is greater than the threshold. When an abnormality or potential fault is detected, a second-level diagnosis is performed. The frequency domain intelligent feature is less than the threshold. If the equipment is deemed to be operating normally, there is no need to proceed to the second level of diagnosis, and the current diagnosis process can be terminated directly. Second-level diagnosis: Input the feature values ​​into the DBN model for diagnosis and obtain the diagnosis results; The DBN model is composed of multiple layers of restricted Boltzmann machines stacked together, and outputs the probability distribution of fault types through a softmax classifier. ; Wherein, the output of the j-th hidden unit for: in, It is the connection weight. and It is a bias term; Update model parameters regularly : in: These are the updated model parameters. It is the adaptive learning rate and the loss function.

[0014] Preferably, the diagnostic results are compressed and transmitted, specifically including the following steps: Predictive differential encoding is performed on the raw time series data of vibration acceleration and sound pressure: in: : The sampled value of the original time series at time n; The predicted value of the sampled value at time n is obtained by linear combination of the previous P historical sampled values; The predicted value for the sampled value at time k is the general form of the prediction model; : The historical sample value at time n; Linear prediction coefficients, representing the weight of the i-th historical sample value to the current predicted value. The Levinson-Durbin recursive algorithm is used to minimize the prediction error. P: Prediction order; Prediction error; Variable-length quantization encoding: for difference sequences Quantization is performed, and a rate-distortion optimization model is established to select the optimal quantization step size. Minimize distortion D at a given bit rate R; Rate-distortion optimized quantization of the difference sequence: in: Rate-distortion cost function, which comprehensively measures the distortion and code rate overhead in the encoding process, is the objective function for optimization; Distortion resulting from quantizing a differential sequence dn at a given bit rate R; R: Encoding bit rate, which is the number of bits required to be transmitted per unit time or per unit of data. Lagrange multipliers are used to balance distortion in rate-distortion optimization. The weights of the bitrate R control the trade-off between compression performance and transmission efficiency; The quantized data is compressed using Varint encoding to obtain the processed data; The processed data and feature parameters are encapsulated into a data frame according to the optimized binary protocol. The frame header includes the sensor ID, high-precision timestamp, data length, feature identifier bit and CRC32 check code. The compressed data stream is organized through an intelligent file block encoding mechanism: the system dynamically adjusts the file block size according to the channel status, uses an enhanced TLV structure for storage, and records the block index, start time, sampling rate metadata and compression algorithm version in the header; the data blocks are temporarily stored in local non-volatile memory to form an adaptive circular buffer; Signal-to-noise ratio (SNR), bit error rate (BER), and round-trip time (RTT) are all mentioned. These are weighting coefficients; when >0.8 Good channel quality: High coding rate LDPC code is used, and transmission efficiency is prioritized; When 0.5 < For channel quality ≤0.8: Reed-Solomon codes are used to balance efficiency and reliability; when ≤0.5 Channel quality is poor: Turbo coding is used, with reliability as the priority; The data payload employs a hybrid coding system based on physical characteristics, first performing predictive differential coding and then variable-length quantization coding. The message header includes a device identifier, message sequence number, time synchronization stamp, and channel status indication. The data is transmitted via a wireless network to a signal relay host deployed on-site. The relay host performs deep analysis, feature fusion, and intelligent aggregation on the data, and uploads the final data in batches to the cloud platform database InfluxDB for long-term storage and trend analysis via an HTTPS RESTful API. After the data is successfully received, the cloud platform returns an ACK confirmation command. Upon receiving the confirmation, the system executes the intelligent shutdown process. When the transmission conditions are met, the optimal transmission strategy is selected based on the real-time LQI evaluation results. After reading the file block from the memory, the data format is restored through a fast decoding process for local display or analysis.

[0015] Preferably, the method further includes the following steps: Dynamic energy budget allocation: The system manages the total energy budget with cloud assistance. And consider energy harvesting ; in, It's about charging efficiency; the system based on... Dynamically adjust the working mode and sampling frequency to ensure continuous operation within the task cycle; PHM-based Remaining Useful Life (RUL) Prediction: In the cloud, using time-series feature data, a performance degradation model is built to predict the remaining useful life of the device. in: Key characteristics reflecting equipment performance degradation (such as vibration energy entropy) are functions of time t; The threshold for determining equipment failure; Failure rate function related to the current working state and environment.

[0016] The present invention also provides an intelligent detection and monitoring system for subways, including a signal transfer control and preprocessing module, an L-shaped support frame and a small portable multimodal intelligent floating plate status sensing terminal, a wideband anti-interference low power consumption multi-mode intelligent track vibration monitoring sensor, a tunnel sidewall multi-parameter vibration and noise intelligent sensing terminal, a support plate, rails and tunnels; The signal relay control and preprocessing module is fixed to the reserved position in the tunnel by bolts or to the temporary position by strong AB glue. The L-shaped support frame is fixed to the tunnel sidewall with strong AB glue; The intelligent sensing terminal for multi-parameter vibration and noise of the tunnel sidewall is attached to the L-shaped support frame by a strong magnetic module at the bottom. The support plate is fixed to the floating plate and the corresponding position below the track with strong AB glue; The powerful magnetic module at the bottom of the small portable multimodal intelligent floating plate status sensing terminal is attached to the support plate. The wideband anti-interference, low-power, multi-mode intelligent track vibration monitoring sensor is fixed to the support plate on the underside of the track by its built-in anchor.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides an intelligent vibration and noise sensor for complex environments such as subways and its data processing method. The sensor achieves high-fidelity acquisition of vibration and noise data, intelligent edge processing, and reliable low-bandwidth transmission through the deep integration of a multimodal sensing architecture and efficient data compression and transmission technology.

[0018] This invention addresses the pain point of balancing power consumption, accuracy, and reliability in existing technologies through a three-layer collaborative optimization architecture of hardware, software, and algorithms, along with a dynamic adaptive learning mechanism. Its core innovations lie in the introduction of three major mechanisms: multi-sensor spatiotemporal fusion sensing, edge intelligent diagnostics based on deep compressed sensing, and reliable transmission with channel adaptation. Attached Figure Description

[0019] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0020] Figure 1 This is a schematic diagram of the arrangement of the present invention; Figure 2 This is a schematic diagram of the sensor's working process according to the present invention; Figure 3 This is a schematic diagram of the data processing flow of the present invention. Detailed Implementation

[0021] Numerous specific details are set forth in the following description to provide a full understanding of the invention. However, the invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0023] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0024] The present invention will now be described in further detail with reference to the accompanying drawings: To better illustrate the technical effects of the present invention, the present invention provides the following specific embodiments to illustrate the above technical process: Example 1: An intelligent detection and monitoring system for subways mainly includes a signal relay control and preprocessing module 1, an L-shaped support frame 2, a small portable multimodal intelligent floating plate status sensing terminal 3, a wideband anti-interference low-power multi-mode intelligent track vibration monitoring sensor 4, a tunnel sidewall multi-parameter vibration and noise intelligent sensing terminal 5, a support plate 6, rails 7, and a tunnel 8. Its layout diagram is shown below. Figure 1 As shown.

[0025] The signal relay control and preprocessing module 1 is fixed to the reserved position in the tunnel with bolts or to the temporary position with strong AB glue. The L-shaped support frame 2 is fixed to the tunnel sidewall with strong AB glue. The tunnel sidewall multi-parameter vibration and noise intelligent sensing terminal 5 is attached to the L-shaped support frame 2 with a strong magnetic module at the bottom. The support plate 6 is fixed to the floating plate and the corresponding position below the track with strong AB glue. The small portable multi-modal intelligent floating plate status sensing terminal 3 is attached to the support plate 6 with a strong magnetic module at the bottom. The wideband anti-interference low power consumption multi-mode intelligent track vibration monitoring sensor 4 is fixed to the support plate 6 on the underside of the track with its own anchor.

[0026] This invention provides an intelligent vibration and noise sensor for complex environments such as subways and its data processing method. The sensor achieves high-fidelity acquisition of vibration and noise data, intelligent edge processing, and reliable low-bandwidth transmission through the deep integration of a multimodal sensing architecture and efficient data compression and transmission technology.

[0027] This invention addresses the pain point of balancing power consumption, accuracy, and reliability in existing technologies by employing a three-layer collaborative optimization architecture of hardware, software, and algorithms, along with a dynamic adaptive learning mechanism. Its core innovations lie in the introduction of three mechanisms: multi-sensor spatiotemporal fusion sensing, edge intelligent diagnostics based on deep compressed sensing, and reliable transmission with channel adaptation.

[0028] 1. System initialization and dynamic awareness wake-up mechanism After the system powers on, a hierarchical initialization protocol is executed to ensure the reliability of the basic support. Regarding the wake-up strategy, a single interrupt-driven approach is abandoned in favor of a three-level dynamic wake-up mechanism.

[0029] Level 1: Hardware-triggered wake-up (extremely low power consumption) The main controller is in deep sleep (Stop) mode, with only the sensor front-end circuitry operating. A built-in programmable comparator continuously monitors the physical signal, generating a hardware interrupt when the signal exceeds a preset digital dynamic threshold, directly waking the main controller. This threshold is not a fixed value but is dynamically adjusted based on the signal noise level. in: Hardware trigger threshold.

[0030] : Baseline average value of sensor signals (calculated during system quiescent periods).

[0031] : Noise standard deviation of sensor signal.

[0032] Configurable sensitivity coefficient.

[0033] Level 2: Software-assisted wake-up (low power) When the sensor reading exceeds a low software warning threshold (such as...) When an anomaly is detected, the system does not immediately start up fully, but instead enters a short-term monitoring mode. In this mode, the main controller collects a small number of samples (e.g., 128) at a low rate for rapid analysis. If an anomaly is confirmed, the system starts up fully; otherwise, it quickly returns to sleep mode.

[0034] Level 3: Adaptive timed inspection (basic power consumption) The system relies on a low-power RTC timer to generate alarm interrupts for periodic checks. The check interval is... It is not fixed, but dynamically adjusted based on the device's historical health status β: in: Basic inspection interval.

[0035] : Maximum adjustment coefficient (0 < \(\alpha\) < 1).

[0036] Equipment health index (derived from historical data analysis; the higher the value, the healthier the equipment).

[0037] Health and safety threshold.

[0038] : Adjust the width parameter of the interval.

[0039] Hyperbolic tangent function, used for smooth transitions.

[0040] This formula enables healthy devices to automatically extend their inspection intervals to save power, while suspected faulty devices have their intervals shortened to enhance monitoring.

[0041] 2. Intelligent Data Processing and Feature Extraction After data acquisition is initiated, spatiotemporal synchronous sampling technology is employed to ensure time alignment of data from multiple sensors. The raw data preprocessing utilizes an improved adaptive Kalman filter algorithm, the core steps of which and its innovations are as follows: State prediction: Covariance prediction: Innovation: Adaptive process noise adjustment Traditional Kalman filtering process noise covariance The present invention makes it adaptively adjust to changes in signal gradient: in: Initial process noise covariance.

[0042] : The gradient (rate of change) of the signal amplitude at time k.

[0043] Smoothing factor.

[0044] When the signal changes drastically Larger, smaller More trust in predicted values ​​to suppress transient interference; when the signal is stable, increase This allows for greater trust in observations and improved tracking accuracy.

[0045] Feature extraction employs a multi-scale deep feature fusion algorithm, surpassing traditional root mean square and kurtosis methods: Temporal enhancement features include root mean square (RMS) and kurtosis. The waveform index (Shape Factor, SI), peak factor (Crest Factor, CF), and impulse factor (Impulse Factor) are also mentioned. )).

[0046] , Frequency domain intelligent features: the spectrum after FFT transformation Perform the calculation.

[0047] Spectrum The calculation method is as follows: For the preprocessed time-domain signal x ( n )( n =0,1,..., L -1, where L Apply a window function to the number of sampling points. w ( n To reduce spectrum leakage: Commonly used window functions include the Hanning window, Hamming window, or Blackman window. For track vibration monitoring, the Hanning window is preferred because it offers a good balance between main lobe width and side lobe suppression. For the windowed signal x w ( n Perform an FFT transform to obtain the complex spectrum. X ( k ): in: k Spectrum line index, corresponding to digital frequency , L : FFT points (usually an integer power of 2, such as 1024, 2048, etc.).

[0048] From the complex spectrum X ( k Extract amplitude information to obtain the amplitude spectrum. A ( k ): in: , is the modulus of the complex spectrum, multiplied by 2 because for real signals, the energy of the negative frequency part folds into the positive frequency part; divided by L is for normalization, so that the amplitude spectrum reflects the true physical amplitude; only k=0 to L / 2 (corresponding to the positive frequency part within the Nyquist frequency) are taken. Convert the spectral line index k to the actual physical frequency fk: Where: fs: sampling frequency (Hz); fk: The actual frequency corresponding to the k-th spectral line.

[0049] After the above steps, the spectrum sequence used for subsequent frequency domain feature calculations is obtained: Where: N=L / 2 (usually removing the 0Hz component corresponding to DC component k=0), and Ak is the amplitude sequence used in frequency domain feature calculation.

[0050] Spectral centroid (FC): Reflects the location where spectral energy is concentrated.

[0051] Spectral variance (FV): reflects the degree of dispersion of the spectrum.

[0052] Upon power-up, the system first initializes the timers, configuring a high-precision real-time clock (RTC) and a watchdog timer to provide a time reference for synchronous sampling by multiple sensors. The sensor units employ a heterogeneous parallel acquisition architecture: the noise sensor unit is based on the high-performance MEMS microphone INMP441, acquiring sound pressure data at a 48kHz sampling rate via the I²S interface; the triaxial vibration sensor unit uses the low-power MEMS accelerometer ADXL1002, synchronously acquiring XYZ triaxial vibration acceleration signals at a maximum sampling rate of 20kHz via a synchronous analog-to-digital converter (ADC). All sensor data is transmitted to the central control module via hardware SPI and DMA channels, achieving high-throughput data acquisition without CPU intervention.

[0053] After the central control module performs real-time preprocessing on the raw data, it executes an enhanced feature extraction algorithm. In addition to calculating the effective value of vibration velocity (RMS), peak acceleration, and equivalent sound level (Leq), it adds the following: Time-frequency domain hybrid features: Extracting wavelet packet energy entropy through wavelet packet transform ( ( ), which can effectively characterize the non-stationary characteristics of a signal.

[0054] , Otherwise, a data frame containing only key feature data is generated. All instructions and data frames are temporarily stored in the local FRAM memory CY15B104Q.

[0055] 3. Intelligent Diagnosis and Decision-Making Mechanism The system employs a two-layer threshold diagnostic model to bridge the gap between "perception" and "cognition." The aforementioned feature values ​​are compared with the two-layer threshold diagnostic model pre-stored in Flash memory.

[0056] The first layer is an adaptive physical threshold: Threshold for the effective value (RMS) of vibration velocity It is no longer a fixed value, but a function related to the frequency f and the ambient temperature T: in It is the threshold under standard conditions. and It is the reference frequency and temperature. and It is a correction factor.

[0057] The frequency domain intelligent feature is greater than the threshold. When an abnormality or potential fault is detected, a second-level diagnosis is performed. The frequency domain intelligent feature is less than the threshold. If the equipment is deemed to be operating normally, there is no need to proceed to the second level of diagnosis, and the current diagnostic process can be terminated directly.

[0058] The second layer employs a Deep Belief Network (DBN) intelligent diagnostic model to diagnose faults and output a fault probability distribution. If the diagnostic result is abnormal, an alarm command in JSON format is generated, including a timestamp, device ID, multi-dimensional feature values, alarm level, and confidence level. Second layer: The extracted 14-dimensional feature vector Input a DBN model. This model consists of multiple layers of Restricted Boltzmann Machines (RBMs) stacked together, and finally outputs the probability distribution of fault types through a softmax classifier. .

[0059] The extracted time-domain enhanced features and frequency-domain intelligent features are combined to form a 14-dimensional feature vector: x 1. Root mean square (derived from the time domain, physical meaning: the overall energy level of the vibration signal). x 2. Kurtosis (derived from the time domain, physical meaning: a measure of the impulse characteristics of a signal, sensitive to early faults). x 3. Waveform Indicators (derived from the time domain, physical meaning: reflecting the smoothness or sharpness of the waveform shape). x 4. Peak index (derived from the time domain, physical meaning: reflects the magnitude of the signal's peak value relative to the effective value). x 5. Impact index (derived from the time domain, physical meaning: reflects the impact intensity of the signal). x 6. Spectral centroid (derived from the frequency domain, physical meaning: reflects the location where spectral energy is concentrated). x 7. Spectral variance (derived from the frequency domain, physical meaning: reflects the degree of dispersion of spectral energy). x 8. Low-frequency band energy ratio (derived from the frequency domain, physical meaning: energy ratio of the 4-50 Hz frequency band). x 9. Energy proportion of the mid-frequency band (derived from the frequency domain, physical meaning: energy proportion of the 50-400 Hz frequency band). x 10 High-frequency band energy ratio (derived from the frequency domain, physical meaning: energy ratio of the 400-1 kHz frequency band). x 11 Dominant frequency amplitude (derived from the frequency domain, physical meaning: the largest amplitude value in the spectrum). x 12 Sideband characteristic coefficients (derived from the frequency domain, physical meaning: reflecting the severity of modulation phenomena). x 13 The magnitude of order 1 (derived from the order spectrum, physical meaning: corresponding to the first-order component of faults such as imbalance). x14 The second-order amplitude (derived from the order spectrum, physical meaning: corresponding to the second-order component of faults such as misalignment).

[0060] The specific components are as follows: Wherein, the output of the j-th hidden unit for: in, It is the connection weight. and It is a bias term.

[0061] The system periodically updates model parameters based on samples verified in the cloud. This enables it to adapt to slow degradation of equipment performance or new failure modes.

[0062] in It is an adaptive learning rate. It is the loss function.

[0063] During the data transmission phase, the intelligent control and data transmission module first activates the low-power wireless module (NB-IoT module BC95), attaches to the network via the AT+CGATT=1 command, and establishes a secure TLS encrypted tunnel with the cloud platform. The transmission protocol employs a channel-adaptive hybrid ARQ mechanism, dynamically selecting the coding scheme based on the real-time calculated Link Quality Index (LQI). 4. High-efficiency data compression and reliable transmission To adapt to low-bandwidth transmission requirements, the system adopts an intelligent two-level coding compression strategy: First, predictive differential coding is performed on the original time-series data of vibration acceleration and sound pressure. , in: : The sampled value of the original time series at time n (such as the actual measured value of vibration acceleration or sound pressure); The predicted value of the sampled value at time n is obtained by linear combination of the previous P historical sampled values; The predicted value for the sampled value at time k is the general form of the prediction model; : The historical sample value at time n; Linear prediction coefficients, representing the weight of the i-th historical sample value to the current predicted value. The Levinson-Durbin recursive algorithm is used to minimize the prediction error. P: Prediction order, which is the number of historical sample values ​​used for prediction; Prediction error (differential signal), which is the difference between the original value and the predicted value, is the core object of data compression.

[0064] Variable-length quantization encoding: for difference sequences Quantization is performed, and a rate-distortion optimization (RD) model is established to select the optimal quantization step size. Minimize distortion D at a given bit rate R.

[0065] The difference sequence is then subjected to rate-distortion optimized quantization: , in: Rate-distortion cost function, which comprehensively measures the distortion and code rate overhead in the encoding process, is the objective function for optimization; Distortion (such as mean square error) that occurs when quantizing a differential sequence dn at a given bit rate R. R: Encoding bit rate, which is the number of bits required to be transmitted per unit time or per unit of data. Lagrange multipliers are used to balance distortion in rate-distortion optimization. The weights of the bitrate R control the trade-off between compression performance and transmission efficiency.

[0066] Finally, Varint encoding is used to compress the quantized data, achieving an overall compression rate of 50-70%.

[0067] The processed data and feature parameters are encapsulated into data frames according to an optimized binary protocol. The frame header includes the sensor ID, high-precision timestamp, data length, feature identifier bits, and CRC32 checksum. The compressed data stream is organized using an intelligent file block encoding mechanism: the system dynamically adjusts the file block size (30-120 seconds) based on channel conditions, uses an enhanced TLV structure for storage, and records the block index, start time, sampling rate metadata, and compression algorithm version in the header. Data blocks are temporarily stored in local non-volatile memory, forming an adaptive circular buffer.

[0068] Signal-to-noise ratio (SNR), bit error rate (BER), and round-trip time (RTT) are all mentioned. Weighting coefficient when >0.8 (Good channel quality): High coding rate LDPC code is used, with transmission efficiency as the priority.

[0069] When 0.5 < ≤0.8 (medium channel quality): Reed-Solomon code is used to balance efficiency and reliability.

[0070] when ≤0.5 (poor channel quality): Use Turbo code, with reliability as the priority.

[0071] The data payload employs a hybrid coding system based on physical characteristics, first performing predictive differential coding followed by variable-length quantization coding, achieving a compression ratio of 50-70%. The message header includes a device identifier, message sequence number, timestamp, and channel status indication. Data is transmitted via wireless network to a signal relay host deployed in the field. The relay host performs deep analysis, feature fusion, and intelligent aggregation on the data, and then uploads the final data in batches to the InfluxDB cloud platform database via an HTTPS RESTful API for long-term storage and trend analysis.

[0072] After successful data reception, the cloud platform returns an ACK confirmation command. Upon receiving the confirmation, the system executes the intelligent shutdown process: the communication module is shut down via the AT+CPWROFF command, the power supply to the external circuits is cut off, and the main controller writes the system status, health index β, and energy budget status to the EEPROM before entering standby mode. The entire system's standby current remains below 1.5 μA until it is woken up again, thus completing a full cycle of optimal performance intelligent monitoring. The sensor workflow diagram and data processing flowchart are shown below. Figure 2 and Figure 3 As shown.

[0073] When the transmission conditions are met, the system selects the optimal transmission strategy based on the real-time LQI evaluation results. After reading the file block from the memory, it restores the data format through a fast decoding process for local display or advanced analysis.

[0074] 5. Cloud-based collaboration and full lifecycle management Dynamic energy budget allocation The system manages the total energy budget with cloud assistance. And consider energy harvesting (vibrational energy). .

[0075] in, It refers to charging efficiency. The system calculates based on... Dynamically adjust the working mode and sampling frequency to ensure continuous operation throughout the task cycle.

[0076] PHM-based Remaining Useful Life (RUL) Prediction. In the cloud, a performance degradation model is built using time-series feature data to predict the remaining useful life of devices.

[0077] in: Key characteristics reflecting equipment performance degradation (such as vibration energy entropy) are functions of time t.

[0078] : The threshold for determining equipment failure.

[0079] Failure rate function related to the current working state and environment.

[0080] The intelligent detection and monitoring method for subways proposed in this invention has the following advantages: 1. Systematic Integration and Functional Synergy of Multi-Source Heterogeneous Sensors: Selection and organic combination of specific functional sensors (SmartVib S041, SmartVN B328-3, etc.). This combination is carefully designed to simultaneously collect physical quantities from different dimensions such as vibration and noise, achieving a leap from "single-point, single-parameter measurement" to "multi-point, multi-parameter fusion and comprehensive diagnosis" of subway conditions. This collaborative work can more comprehensively and accurately depict the overall situation of track health, train operation status, and environmental noise, avoiding information silos.

[0081] 2. Edge Intelligent Data Processing Paradigm: This paradigm shifts computing power to the signal relay module, upgrading it from a simple data repeater to an edge intelligent node. It enables local real-time data processing (filtering, feature extraction, anomaly detection algorithms), allowing for immediate assessment of device status and triggering alerts.

[0082] 3. High-Reliability Hybrid Networking and Data Relay Architecture: A layered hybrid transmission network architecture was designed. The front-end sensing layer uses low-power wide-area network (LPWAN) technologies (Varint, ZigBee) to connect sensors and relay modules, solving the power supply and cabling problems in tunnels; the back-end network layer utilizes high-speed and reliable links (5G / Industrial Ethernet) for remote backhaul.

[0083] 4. Ultra-low power system architecture based on dynamic perception and intelligent decision-making This protection point focuses on the collaborative operation of the entire system, aiming to solve the fundamental problem of "how to achieve intelligent monitoring while maintaining extremely low power consumption." Its specific protection measures include: (1) Three-level dynamic wake-up mechanism: the specific implementation methods, collaborative working logic of the three modes of hardware triggering, software assistance, and adaptive timing, and the dynamic threshold algorithm behind them ( ) and the adaptive sleep interval algorithm (the formula for calculating Ts).

[0084] (2) Graded initialization and intelligent shutdown process: including technologies such as temperature compensation RTC and heartbeat watchdog to ensure basic reliability, as well as intelligent state saving and shutdown process based on system status and energy budget.

[0085] (3) Hardware-Software-Algorithm Three-Layer Collaborative Optimization Architecture: It is emphasized that these three are not simply superimposed, but deeply coupled, together forming a systematic solution to achieve the power consumption and performance indicators.

[0086] 5. Multi-dimensional feature extraction and fusion decision-making method for edge intelligent diagnosis This protection point focuses on the data-to-information transformation process, aiming to solve the problem of "how to achieve accurate fault detection at resource-constrained edge environments." Its specific protection content includes: (1) Improved adaptive Kalman filter algorithm: especially its process noise covariance Qk varies with signal gradient The specific model of adaptive adjustment ( ).

[0087] (2) Multi-scale deep feature fusion algorithm: explicitly protects the specific feature combinations extracted, including but not limited to: Temporal enhancement features: Calculation methods for kurtosis (κ) and impact index (Iim).

[0088] Frequency domain intelligent features: formulas for calculating the spectral centroid (FC) and spectral variance (FV).

[0089] Time-frequency domain characteristics: Calculation method of wavelet packet energy entropy (Hw).

[0090] (3) Two-layer threshold diagnostic model: First layer: Adaptive physical threshold model ( ).

[0091] The second layer: the structure of the optimized deep belief network (DBN) diagnostic model applied to the edge and its incremental learning mechanism (the update formula of θnew).

[0092] 6. High-efficiency data compression and transmission protocol based on channel awareness and physical characteristics This protection point focuses on the efficient and reliable transmission of information, aiming to solve the problem of "how to ensure complete data transmission without excessive power consumption in harsh wireless environments." Its specific protection features include: (1) Hybrid coding method based on physical features: The specific implementation of predictive differential coding includes the calculation of the predicted value ân and the solution method for the prediction coefficient φi.

[0093] Combination rate distortion optimization model ( , A variable-length quantization encoding method is proposed to find the optimal quantization step size. .

[0094] (2) Channel-adaptive reliable transmission protocol: A comprehensive evaluation model for Link Quality Index (LQI) ).

[0095] The decision-making mechanism based on LQI enables real-time and dynamic switching between different error correction coding schemes such as LDPC, Reed-Solomon, and Turbo.

[0096] (3) Intelligent file block encoding and adaptive ring buffer: a mechanism for dynamically adjusting the file block size according to the channel state, and an enhanced TLV storage structure to cooperate with it.

[0097] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules, units, or units is merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units, modules, or components may be combined or integrated into another device, or some features may be ignored or not executed.

[0098] The units may or may not be physically separate. The components shown as units can be one or more physical units, meaning they can be located in one place or distributed in multiple different locations. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0099] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0100] In particular, according to embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit (CPU), it performs the functions defined in the methods of this invention. It should be noted that the computer-readable medium described above in this invention can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof.

[0101] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0102] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions within the technical scope disclosed in the present invention should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A smart detection and monitoring method for subways, characterized in that, Includes the following steps: Acquire sensor signals; The sensor signal is preprocessed to obtain the preprocessed sensor signal; Feature extraction is performed on the preprocessed sensor signal to obtain feature values; The diagnostic results are obtained by comparing the feature values ​​with the two-layer threshold. The diagnostic results are compressed and transmitted.

2. The intelligent detection and monitoring method for subways according to claim 1, characterized in that, Acquiring sensor signals specifically includes the following steps: The sensor signals are continuously monitored while the main controller is in deep sleep mode; When the sensor signal exceeds the dynamic threshold, the main controller is activated. After waking up the main controller, the sensor signals are analyzed, and a full start is initiated if the analysis result is abnormal. After full operation, regular inspections will be conducted.

3. The intelligent detection and monitoring method for subways according to claim 2, characterized in that, The formula for calculating the dynamic threshold is: in: For dynamic thresholds; This represents the baseline average value of the sensor signal; The noise standard deviation of the sensor signal; This is a configurable sensitivity coefficient; The formula for calculating the warning threshold is: in: This is the warning threshold; The inspection interval of the timed inspection The calculation formula is: in: Basic inspection interval; This is the maximum adjustment range coefficient; Equipment health index; For health and safety thresholds; To adjust the width parameter of the interval; It is a hyperbolic tangent function used for smooth transitions.

4. The intelligent detection and monitoring method for subways according to claim 3, characterized in that, The sensor signal is preprocessed to obtain the preprocessed sensor signal. The specific steps include: An improved adaptive Kalman filter algorithm is used to preprocess the sensor signal: State prediction: Covariance prediction: Process noise covariance Adaptive adjustment as the signal gradient changes: in: The initial process noise covariance; Let k be the gradient of the signal amplitude at time k; It is a smoothing factor; When the signal changes drastically, reduce More trust in predicted values ​​to suppress transient interference; when the signal is stable, increase This allows for greater trust in observations and improved tracking accuracy.

5. The intelligent detection and monitoring method for subways according to claim 4, characterized in that, Feature extraction is performed on the preprocessed sensor signal to obtain feature values, specifically including the following steps: A multi-scale deep feature fusion algorithm is used to extract features from the preprocessed sensor signal: Temporal augmentation features: including root mean square (RMS) and kurtosis. Waveform index SI, peak index and impact index ; ， Frequency domain intelligent features: the spectrum after FFT transformation Perform calculations; Spectral centroid FC: reflects the location where spectral energy is concentrated; Spectral variance FV: reflects the degree of dispersion of the spectrum; After the central control module performs real-time preprocessing on the raw data, it executes an enhanced feature extraction algorithm. In addition to calculating the effective value of vibration velocity, peak acceleration, and equivalent sound level, it adds the following: Time-frequency domain hybrid features: Extracting wavelet packet energy entropy through wavelet packet transform Characterizing the non-stationary properties of a signal: ， 。 6. The intelligent detection and monitoring method for subways according to claim 5, characterized in that, The spectrum The calculation method is as follows: For the preprocessed time-domain signal x(n), n=0,1,...,L-1, where L is the number of sampling points; apply a window function w(n) to reduce spectral leakage: The Hanning window is chosen as the window function: Performing an FFT on the windowed signal xw(n) yields the complex spectrum X(k): Where: k is the spectrum line index, corresponding to the digital frequency. L is the number of FFT points; The amplitude information is extracted from the complex spectrum X(k) to obtain the amplitude spectrum A(k): in: , where is the modulus of the complex spectrum; Convert the spectral line index k to the actual physical frequency fk: Where: fs is the sampling frequency; fk is the actual frequency corresponding to the kth spectral line; Spectral sequence calculated from frequency domain features: Where N = L / 2, and Ak is the amplitude sequence used in the frequency domain feature calculation.

7. The intelligent detection and monitoring method for subways according to claim 6, characterized in that, The diagnostic results are obtained by comparing the feature values ​​with a two-layer threshold, specifically including the following steps: Compare the feature values ​​with the adaptive physical threshold; The adaptive physical threshold for: in: It is the threshold under standard conditions. and It is the reference frequency and temperature. and It is a correction factor; The frequency domain intelligent feature is greater than the threshold. When an abnormality or potential fault is detected, a second-level diagnosis is performed. The frequency domain intelligent feature is less than the threshold. If the equipment is deemed to be operating normally, there is no need to proceed to the second level of diagnosis, and the current diagnosis process can be terminated directly. Second-level diagnosis: Input the feature values ​​into the DBN model for diagnosis and obtain the diagnosis results; The DBN model is composed of multiple layers of restricted Boltzmann machines stacked together, and outputs the probability distribution of fault types through a softmax classifier. ; Wherein, the output of the j-th hidden unit for: in, It is the connection weight. and It is a bias term; Update model parameters regularly : in: These are the updated model parameters. It is the adaptive learning rate and the loss function.

8. The intelligent detection and monitoring method for subways according to claim 7, characterized in that, The process of compressing and transmitting diagnostic results includes the following steps: Predictive differential encoding is performed on the raw time series data of vibration acceleration and sound pressure: in: : The sampled value of the original time series at time n; The predicted value of the sampled value at time n is obtained by linear combination of the previous P historical sampled values; The predicted value for the sampled value at time k is the general form of the prediction model; : The historical sample value at time n; Linear prediction coefficients, representing the weight of the i-th historical sample value to the current predicted value. The Levinson-Durbin recursive algorithm is used to minimize the prediction error. P: Prediction order; Prediction error; Variable-length quantization encoding: for difference sequences Quantization is performed, and a rate-distortion optimization model is established to select the optimal quantization step size. Minimize distortion D at a given bit rate R; Rate-distortion optimized quantization of the difference sequence: in: Rate-distortion cost function, which comprehensively measures the distortion and code rate overhead in the encoding process, is the objective function for optimization; Distortion resulting from quantizing a differential sequence dn at a given bit rate R; R: Encoding bit rate, which is the number of bits required to be transmitted per unit time or per unit of data. Lagrange multipliers are used to balance distortion in rate-distortion optimization. The weights of the bitrate R control the trade-off between compression performance and transmission efficiency; The quantized data is compressed using Varint encoding to obtain the processed data; The processed data and feature parameters are encapsulated into a data frame according to the optimized binary protocol. The frame header includes the sensor ID, high-precision timestamp, data length, feature identifier bit and CRC32 check code. The compressed data stream is organized through an intelligent file block encoding mechanism: the system dynamically adjusts the file block size according to the channel status, uses an enhanced TLV structure for storage, and records the block index, start time, sampling rate metadata and compression algorithm version in the header; the data blocks are temporarily stored in local non-volatile memory to form an adaptive circular buffer; Signal-to-noise ratio (SNR), bit error rate (BER), and round-trip time (RTT) are all mentioned. These are weighting coefficients; when >0.8 Good channel quality: High coding rate LDPC code is used, and transmission efficiency is prioritized; When 0.5 < For channel quality ≤0.8: Reed-Solomon codes are used to balance efficiency and reliability; when ≤0.5 Channel quality is poor: Turbo coding is used, with reliability as the priority; The data payload employs hybrid coding based on physical characteristics, first performing predictive differential coding and then variable-length quantization coding; the message header includes device identifier, message sequence number, time synchronization stamp, and channel status indication; the data is sent via wireless network to a signal relay host deployed on-site, where the relay host performs deep analysis, feature fusion, and intelligent aggregation on the data, and uploads the final data in batches to the cloud platform database InfluxDB for long-term storage and trend analysis via HTTPS RESTful API; After the data is successfully received, the cloud platform returns an ACK confirmation command. Upon receiving the confirmation, the system executes the intelligent shutdown process. When the transmission conditions are met, the optimal transmission strategy is selected based on the real-time LQI evaluation results. After reading the file block from the memory, the data format is restored through a fast decoding process for local display or analysis.

9. The intelligent detection and monitoring method for subways according to claim 8, characterized in that, It also includes the following steps: Dynamic energy budget allocation: The system manages the total energy budget with cloud assistance. And consider energy harvesting ; in, It's about charging efficiency; the system based on... Dynamically adjust the working mode and sampling frequency to ensure continuous operation within the task cycle; PHM-based remaining useful life prediction: In the cloud, using time-series feature data, a performance degradation model is built to predict the remaining useful life of the equipment. in: The key characteristic reflecting equipment performance degradation is a function of time t; The threshold for determining equipment failure; Failure rate function related to the current working state and environment.

10. A smart detection and monitoring system for subways, used to implement the smart detection and monitoring method for subways as described in any one of claims 1-9, characterized in that, Includes a signal relay control and preprocessing module, an L-shaped support frame and a small portable multimodal intelligent floating plate status sensing terminal, a wideband anti-interference low-power multi-mode intelligent track vibration monitoring sensor, a tunnel sidewall multi-parameter vibration and noise intelligent sensing terminal, support plate, rail and tunnel; The signal relay control and preprocessing module is fixed to the reserved position in the tunnel by bolts or to the temporary position by strong AB glue. The L-shaped support frame is fixed to the tunnel sidewall with strong AB glue; The intelligent sensing terminal for multi-parameter vibration and noise of the tunnel sidewall is attached to the L-shaped support frame by a strong magnetic module at the bottom. The support plate is fixed to the floating plate and the corresponding position below the track with strong AB glue; The powerful magnetic module at the bottom of the small portable multimodal intelligent floating plate status sensing terminal is attached to the support plate. The wideband anti-interference, low-power, multi-mode intelligent track vibration monitoring sensor is fixed to the support plate on the underside of the track by its built-in anchor.

Images