Intelligent polishing state monitoring method based on multi-sensor fusion

Abstract

The application discloses an intelligent polishing state monitoring method based on multi-sensor fusion, and relates to the technical field of industrial automation and intelligent control.The method is based on multi-source features of grinding specific energy, thermal shock, mechanical vibration and surface morphology to construct a comprehensive state value to represent the overall load level;by analyzing the time sequence evolution of the comprehensive state value, the prediction residual is extracted and a dynamic noise benchmark is established, and then the interference degree of the current environment is evaluated;based on the interference degree, a global scaling factor is generated to synchronously and adaptively correct the discrimination benchmark of each feature branch, and the confidence is calculated based on the abnormal grid point proportion, the comprehensive state value and the adaptive benchmark value;according to the confidence level, multi-dimensional collaborative verification is performed, and the corresponding hierarchical response strategy is triggered.The application realizes the intelligent monitoring of the whole process of the polishing operation from state perception, reliable evaluation to control through the adaptive adjustment mechanism, and improves the intelligent level of the polishing process monitoring.

Description

Technical Field

[0001] This invention relates to the field of industrial automation and intelligent control technology, and to an intelligent grinding status monitoring method based on multi-sensor fusion. Background Technology

[0002] Rail grinding is a critical process in railway maintenance, and its condition directly determines the surface quality of the rails and operational safety. Current methods for monitoring rail grinding operations primarily rely on fixed threshold alarms from single sensors such as vibration and temperature sensors. However, grinding operations are complex and variable; factors such as rail material, grinding wheel condition, and cutting depth can significantly alter the normal range of sensor readings. Fixed thresholds either frequently trigger false alarms during heavy cutting or fail to detect real risks during light cutting. Furthermore, non-grinding interferences such as oil stains, water films, and ballast splashes frequently occur at the grinding site. Traditional statistical filtering cannot distinguish between genuine grinding anomalies and false sensor interference, and contaminated data easily leads to misjudgments. While recent research has attempted multi-source data fusion, most methods employ fixed weights or neural networks trained on historical data, lacking physical mechanism support and failing to adapt to the variable operating conditions. More importantly, current technology stops at alarms, failing to form a closed loop of monitoring-decision-execution. The system lacks self-adjustment capabilities, is sensitive to occasional interference, and has a high false alarm rate, severely impacting grinding efficiency and operational continuity.

[0003] In summary, existing technologies, when dealing with complex and ever-changing grinding conditions, are limited by the rigidity of fixed thresholds, the lack of interference recognition capabilities, and the lack of fusion strategies supported by physical mechanisms. This results in a high false alarm rate, poor environmental adaptability, and an inability to achieve closed-loop control of the work status.

[0004] Therefore, there is an urgent need for an intelligent grinding status monitoring method that can automatically remove invalid data interference, has the ability to adapt to working conditions, and improves the robustness of fault identification through multi-dimensional collaborative analysis, so as to improve the intelligence level and reliability of grinding operations. Summary of the Invention

[0005] This application provides an intelligent grinding status monitoring method based on multi-sensor fusion, which aims to solve the technical problems in the prior art, such as poor environmental adaptability caused by rigid monitoring standards, difficulty in effectively removing non-steady-state interference, and insufficient robustness of fault identification.

[0006] This application provides a method for intelligent grinding status monitoring based on multi-sensor fusion, including: The system acquires multidimensional heterogeneous raw data reflecting the grinding state, geometric change features reflecting the material removal effect, and mileage coding signals in real time during the grinding operation. Based on the mileage coding signals, a unified spatial reference is constructed, and the multidimensional heterogeneous raw data and geometric change features are mapped to the spatial reference to obtain a spatiotemporally aligned dataset. The multidimensional heterogeneous raw data includes at least three-dimensional grinding force, temperature field distribution image, vibration acceleration, and visual image of the grinding area. The geometric change features include at least real-time cutting depth and real-time grinding width. Grinding drive features and multidimensional response features are extracted from each mileage grid point in the spatiotemporal aligned dataset, and real-time grinding specific energy is calculated based on the mapping relationship between the two. If any of the multidimensional response features and real-time grinding specific energy of the current mileage grid point deviates from its corresponding expected response allowable range, the grid point is marked as an anomaly. According to the anomaly marking result, the data in the anomaly grid point is removed, and the data of the remaining mileage grid points are reorganized into a valid feature sequence. The multidimensional response features include at least the equivalent grinding force value, temperature rise value, and spectral energy value, and the grinding drive features include at least the statistical features of geometric change features. An iterative processing cycle is constructed, and the adaptive reference value of each feature branch in the previous cycle is obtained. Based on the adaptive reference value, the deviation feature of the feature value of each feature branch in the current cycle relative to the adaptive reference value is determined. The grinding specific energy distribution feature in this cycle is used as a global prior constraint. Combined with the deviation feature, the weight of each feature branch in this cycle is dynamically adjusted. The result of the weighted fusion of the feature values ​​of each feature branch is represented as the comprehensive state value of this cycle. The iterative processing cycle is composed of several continuous processing units based on the data distribution characteristics of the effective feature sequence. The feature value of each feature branch represents the transient thermal shock degree, mechanical vibration instability trend and abnormal surface morphology state in the current cycle. Based on the comprehensive state value and its historical evolution trend of the current iteration processing cycle, an environmental noise observer is constructed to calculate the noise benchmark that characterizes the degree of environmental interference in real time. The noise benchmark is used to adaptively correct the benchmark values ​​of each branch feature value of the previous iteration processing cycle to obtain the adaptive benchmark values ​​of each feature branch of the current iteration processing cycle, which are used for the next iteration processing cycle. Within the current iteration cycle, multi-dimensional collaborative verification is performed based on the proportion of abnormal grid points, the comprehensive state value, and the evolution characteristics of the adaptive baseline value determined by the abnormal marking results. According to the working condition confidence level generated by the verification results, a hierarchical response strategy is triggered to realize closed-loop intelligent monitoring of grinding operations.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: (1) Data cleaning based on grinding specific energy and multi-physical quantity constraints removes invalid data caused by air cutting, slippage or sensor noise from the original data, ensuring that subsequent fusion calculation and condition assessment are based only on the effective area where grinding action actually occurs, avoiding interference of invalid data with monitoring results and improving the data effectiveness of condition monitoring.

[0008] (2) By introducing grinding specific energy as a weight adjustment factor, the adaptive fusion weights of the three branches of thermal shock, mechanical vibration and surface morphology are dynamically constructed, so that the comprehensive state value can automatically adjust the sensitivity to thermal, vibration and morphological characteristics according to the current energy load level, thereby generating a comprehensive state value that can fully characterize the overall health of the current grinding condition, and realizing the complementarity and enhancement of multi-source features.

[0009] (3) By monitoring the time-series fluctuations of the comprehensive state value, calculating the prediction residual and relative disturbance amplitude, the current environmental background noise level is deduced, and then a global scaling factor is constructed to update the benchmark values ​​of the three branches of thermal shock, mechanical vibration and surface morphology. When the environmental noise is high, the benchmark is automatically raised to avoid false alarms, and when the environment is stable, the benchmark is automatically lowered to improve sensitivity, thus realizing the real-time matching between the monitoring standard and the current working condition.

[0010] (4) When performing adaptive verification, a multi-dimensional verification system was constructed, which includes the proportion of abnormal grid points, the overall state value and the changing trend of dynamic benchmark value. The proportion of abnormal grid points is used to determine the scope of the fault, and the overall state value is used to determine the severity of the fault, so as to achieve mutual verification between space and time. The changing trend of dynamic benchmark value is introduced as a reference system, and its synergy with the overall state value is used to distinguish the authenticity of the fault, so as to achieve the synergy between signal and benchmark. The multi-dimensional verification system greatly improves the system's ability to identify real faults under complex working conditions, effectively eliminates local interference and random noise, significantly improves the fault identification accuracy under complex working conditions, and ensures the accuracy of the hierarchical response strategy.

[0011] (5) Based on the hierarchical response strategy of multi-index fusion, it can provide different levels of processing suggestions or control instructions according to the severity of the fault, avoiding a one-size-fits-all shutdown process. While ensuring processing quality, it maximizes the continuity and efficiency of grinding operations and improves the intelligence level of grinding process monitoring.

[0012] In summary, this invention proposes for the first time an adaptive monitoring paradigm that drives benchmark evolution through state evolution. By deduce the temporal fluctuation characteristics of the comprehensive state value, the dynamic noise benchmark under the current working condition is inferred in real time, and the discrimination benchmark of all sensing branches is updated synchronously accordingly. This realizes the automatic adjustment of the judgment standard with the evolution of working conditions and solves the technical problems of poor adaptability and high false alarm rate of traditional methods under changing working conditions. Attached Figure Description

[0013] Figure 1 This is a schematic flowchart illustrating an intelligent grinding status monitoring method based on multi-sensor fusion, provided as an embodiment of this application. Detailed Implementation This application provides a method for intelligent grinding status monitoring based on multi-sensor fusion, such as... Figure 1 As shown, the method specifically includes the following steps: S100: Real-time acquisition of multidimensional heterogeneous raw data reflecting the grinding state, geometric change features reflecting the material removal effect, and mileage coding signals during the grinding operation. Based on the mileage coding signals, a unified spatial reference is constructed, and the multidimensional heterogeneous raw data and geometric change features are mapped to the spatial reference to obtain a spatiotemporally aligned dataset. The multidimensional heterogeneous raw data includes at least three-dimensional grinding force, temperature field distribution image, vibration acceleration, and visual image of the grinding area. The geometric change features include at least real-time cutting depth and real-time grinding width.

[0014] Specifically, high-frequency force sensors, infrared thermal imagers, laser profilometers, triaxial accelerometers, and odometer encoders are installed on the grinding equipment, which is typically a grinding cart. This embodiment employs a tiered layout strategy encompassing pre-testing, mid-grinding, and post-verification. Taking the contact point between the grinding wheel and the rail as the origin, the sensors are arranged sequentially along the grinding cart's direction of travel: First tier: The pre-scanning laser profilometer, located at the very front of the cart, is responsible for reconnaissance of the original terrain, establishing a baseline map, and completely avoiding operational interference. Second tier: The grinding motor assembly in the middle, integrating force sensors and accelerometers, performs the cutting action and captures the original cutting dynamics. Third tier: The infrared thermal imager, located immediately behind the work area, avoids direct sparks and captures the temperature field before heat dissipation. Fourth tier: The industrial camera, positioned slightly behind, uses strobe-freezing to avoid the effects of high-temperature heat radiation and clearly captures surface texture. Fifth tier: The post-scanning laser profilometer, located at the very rear, performs the highest-precision geometric dimension acceptance measurement after the rail surface has completely cooled and stabilized. Specifically, a high-frequency force sensor is embedded in the flange connection between the grinding motor spindle and the grinding wheel, or directly integrated into the key stress point of the grinding head suspension mechanism, to collect three-dimensional grinding force; an infrared thermal imager is fixed at an angle of approximately 30°~45° above and behind the grinding head, with the lens facing the surface of the steel rail that has just been ground, lagging behind the grinding by 100-200mm; the laser profilometer system adopts a dual-probe collaborative layout: two high-precision laser profilometers are configured, defined as a pre-scanning profilometer and a post-scanning profilometer, respectively. The pre-scanning laser profilometer is fixedly installed at the very front of the grinding carriage, at least 2-5 meters in front of the grinding head's direction of travel, ensuring that its scanning area is completely in front of all grinding wheel contact points. The rear-scanning laser profilometer is mounted on the detection beam behind the grinding head, projecting a laser line vertically downwards to the top and side surfaces of the rail. It is located after the infrared thermal imager, lagging behind the grinding point by approximately 400-600mm. A triaxial accelerometer is rigidly fixed to the grinding motor housing or the grinding head suspension arm at the point closest to the vibration source. An industrial vision sensor, equipped with a high-intensity stroboscopic light source, is mounted directly behind the grinding head, looking vertically down or obliquely at the rail surface, located between or adjacent to the rear-scanning laser profilometer. A mileage encoder is mounted on the axle end of the non-powered running wheelset (driven wheelset) at the bottom of the grinding carriage, preferably on the axle end of a dedicated measuring carriage bogie, or on the axle end of any non-driven driven wheel of the grinding head carriage.

[0015] Furthermore, the spatiotemporal aligned dataset is obtained, including: During the operation of the grinding vehicle, a pre-scanning laser profiler mounted at the front of the rail grinding equipment passes through the current section to be ground before the grinding head, and collects and generates an initial profile data sequence of the rail before grinding. The initial profile data sequence consists of several sampling points distributed along the transverse direction of the rail. After grinding is completed, the multi-source sensor detection unit mounted on the back end of the rail grinding equipment is used to collect data on the ground section in real time, generating a rail profile data sequence after grinding, a three-dimensional grinding force signal and high-frequency vibration acceleration signal time sequence, a temperature field distribution image and a visual image time sequence of the grinding area. Meanwhile, the mileage encoder integrated into the walking mechanism of the grinding equipment synchronously records the mileage pulse signal. Using the mileage pulse signal, a continuous actual mileage coordinate axis along the extension direction of the rail is constructed as a unified spatial reference. Based on the spatial reference, the data sequence of the polished rail profile is spatially aligned with the initial profile data sequence, and the corresponding values ​​at the same mileage coordinates are extracted for comparison and calculation to generate a time series of geometric change features characterizing the material removal effect. The geometric change features include at least real-time cutting depth and real-time grinding width. Based on the preset spatial resolution, a standard mileage grid with equal spacing is generated on the continuous actual mileage coordinate axis. The time series of geometric change features, three-dimensional grinding force signal, high-frequency vibration acceleration signal, temperature field distribution image and visual image time series are mapped to the corresponding standard mileage grid, and the data at each grid point are aggregated to form a spatiotemporally aligned dataset.

[0016] Specifically, the dynamic acquisition of the initial baseline profile of the rail involves the pre-scanning laser profilometer passing through the section to be processed first as the grinding vehicle moves at its operating speed. Since this position completely avoids the sparks, high-temperature heat radiation, and mechanical vibrations generated during grinding, the sensor can acquire the original cross-sectional point cloud of the rail with an extremely high signal-to-noise ratio. The system marks and temporarily stores this data in a ring buffer based on mileage index, serving as the absolute geometric reference for subsequent calculations of material removal. Synchronous triggering of the multi-sensor array in the working area: as the vehicle continues to move and the grinding head completes its cutting action, the post-scanning laser profilometer acquires the instantaneous cross-sectional point cloud of the ground surface; high-frequency force sensors and accelerometers continuously output three-dimensional grinding force signals and vibration acceleration signals at a frequency of ≥5000Hz to capture transient impact and resonance characteristics during the cutting process; an infrared thermal imager outputs a temperature field matrix at a frame rate of ≥60fph, forming a high-resolution temperature field distribution image time series and recording the diffusion trajectory of grinding heat; an industrial vision sensor, in conjunction with a high-intensity stroboscopic light source, freezes high-speed motion images within millisecond-level exposure time, acquiring high-resolution visual images of the rail surface. The odometer encoder outputs high-precision pulse signals in real time, serving as the time-space anchor point for the entire system. The laser profilometer scans the cross-section of the rail, which is a section perpendicular to the rail's extension direction. The data points on this section are distributed transversely along the rail.

[0017] Discrete mileage points are calculated by accumulating discrete pulse signals output from the onboard mileage encoder and combining them with wheel diameter parameters. Subsequently, linear interpolation or spline interpolation methods are used to smooth the transitions between adjacent discrete mileage points, thus constructing a continuous actual mileage coordinate axis S along the rail extension direction. This coordinate axis serves as a unified spatial reference for all subsequent profile data acquisition, buffering, and matching. First, a mapping relationship between the sensor's time domain and the rail's spatial domain is established. The mileage encoder outputs discrete position signals with a fixed physical resolution, such as each pulse representing 0.5 mm. Real-time buffered pulse sequence {(t k ,S k )}, where T k S is the system time when the k-th pulse is generated. k S represents the absolute mileage accumulated at that moment. k =k·P res P res The encoder's single-pulse resolution is given. For any sensor data packet, its precise acquisition timestamp t is recorded. i , due to t i It usually falls between two adjacent pulse times (t) k-1 ≤t i ≤t k Furthermore, the operating speed of the grinding equipment within a millisecond-level time interval can be considered to change linearly and uniformly. The system uses a linear time interpolation algorithm to calculate t. i The corresponding absolute track mileage coordinates S i S i =S k-1 +λ·(S k -S k-1 ), Where λ is the time normalization coefficient, with a value range of [0,1], and S k-1 and S k These are the cumulative mileage values ​​corresponding to the (k-1)th and kth pulses, respectively. All heterogeneous sensor data are assigned a unique spatial label S. i This completes the first transformation from the discrete time domain to the continuous spatial domain. To achieve unified alignment of multi-source data, the system sets a unified standard mileage grid step size ΔS, where ΔS = 1mm or 5mm, based on the grinding process accuracy requirements and the spatial resolution of the slowest sensor. A segmented fixed window strategy is used to construct the mileage grid. The system divides continuous grinding operation segments into several standard processing units (windows) of fixed length, and within each unit, the total number of grids M remains constant. Based on the memory capacity and real-time requirements of the vehicle control system, a standard processing window length L is preset. win For example, 100m or 500m, the starting mileage S startFor the first window, the mileage reading is taken from the first valid profile data point of the working section captured by the front laser profilometer. For subsequent windows, it is automatically updated to the ending mileage S of the previous window. start (K)=S end (K-1) enables seamless window transitions. Termination mileage S end S is calculated from the initial mileage and the fixed window length. end =S start +L win If the remaining length at the actual end of the operation is less than L win If the remaining portion is treated as the last incomplete window, then S end Retrieve the actual destination mileage. Based on the starting mileage S of the current window. start and the final mileage S end The system generates a series of discrete, equally spaced standard mileage reference points, denoted as: Where j is the grid index, a positive integer representing the j-th data node in the sequence, corresponding to the j-th virtual scale position on the rail. Here, represents the absolute coordinates of the grid, indicating the absolute mileage value of the j-th grid node. M represents the total number of grid points generated within the entire section, determined by the total grinding length L = S. end -S start The decision is made, and the calculation formula is as follows: , Indicates rounding down. Sequence This forms the spatial skeleton of the spatiotemporally aligned dataset, with real-time acquired multi-source sensor data mapped onto the grid sequence of the current window. middle.

[0018] To unify data collected from different sensors with varying time sampling rates onto a common spatial reference for comparative analysis, this embodiment first constructs a virtual, equally spaced spatial reference coordinate system, called the spatial skeleton. This skeleton consists of a series of discrete virtual scale positions distributed at fixed mileage intervals. These virtual scale positions are a series of standardized spatial nodes evenly distributed along the rail's extension direction. These nodes are not actual physical markers on the rail, such as welds or sleepers, but rather virtual ruler scales established by the data processing system to unify the spatial reference of multi-source data. Each virtual scale position corresponds to a theoretical cross-sectional position on the rail, used to carry all sensor data collected in the vicinity of that position. The system iterates through the data already assigned spatial labels S. i The set of all original sampling points {(t i S i V i )}, V iThis represents the raw sensor reading. For each raw sampling point S... i The standard grid index j to which it belongs is directly calculated through the floor operation. Boundary validity determination: If the calculated value is 1≤j≤M, then the point is determined to belong to the current processing window, and subsequent data aggregation steps are executed. If j>M or j<1, then the point is determined to belong to the next window or is invalid data, temporarily stored in the cache for processing in the next iteration cycle, or directly discarded. start This represents the starting mileage of the current processing window. The complete information of the original sampling points {(t)} i S i V i The data is directly appended to the original data list corresponding to the j-th grid, completing the data bucketing process. The temperature field distribution image is associated with the discrete image data of the visual image to obtain the mileage coordinates S corresponding to the exposure center time of each frame. ing Similarly, using the formula The system determines the primary grid index j to which the image frame belongs and stores the original image file's storage path, frame ID, exposure timestamp, and image resolution parameters in this list. If multiple images fall into the same grid j due to extremely slow vehicle speed, the original data list will contain multiple image indices, arranged chronologically. If extremely fast vehicle speed results in no images falling into a certain grid, the list will be empty or marked as having no data, and no image compositing or interpolation will be performed. After completing the above aggregation and bucketing, the system generates the final spatiotemporally aligned dataset. This dataset consists of a series of data units arranged by mileage index j, with a spatial index.

[0019] Furthermore, a time series of geometrical change features characterizing the material removal effect is generated, including: The real-time operating speed of the rail grinding equipment and the physical installation distance between the pre-scanning laser profilometer and the post-scanning laser profilometer are acquired simultaneously. The time delay compensation is calculated in real time. Time delay compensation = physical installation distance / real-time operating speed. Based on the real-time running speed and mileage accumulation algorithm, the sequence of polished rail profile data collected at time t is mapped to the current mileage coordinate on the continuous actual mileage coordinate axis to form a real-time profile data sequence. Simultaneously, by using the time delay compensation amount to backtrack and calculate, the historical acquisition time corresponding to the current mileage coordinates is obtained, and the initial profile data sequence of the rail before grinding recorded at the historical acquisition time is extracted from the cache, and associated with the current mileage coordinates to form an initial profile data sequence, thereby achieving spatiotemporal alignment under a unified mileage spatial benchmark. On the continuous actual mileage coordinate axis, the initial profile value and real-time profile value corresponding to the same mileage coordinate are obtained to form a same mileage paired dataset. Based on the same mileage paired dataset, the following feature calculation is performed: Calculate the real-time cutting depth: Based on the lateral coordinates of each sampling point in the paired dataset, align the initial profile value and the real-time profile value laterally to obtain corresponding point pairs at the same lateral position, calculate the distance difference between each corresponding point in the vertical direction of the rail, form a distance difference sequence, remove the maximum value of the preset proportion in the sequence, select the maximum value in the remaining sequence, or use the high-order preset quantile value of the sequence as the real-time cutting depth of the rail cross section at the current mileage coordinate; Calculate the real-time grinding width: filter out sampling points whose absolute value of the distance difference exceeds the preset grinding threshold, and take them as valid sampling points. Aggregate the horizontally adjacent valid sampling points into at least one continuous grinding area. Select the main grinding area with the largest horizontal span. Based on the horizontal coordinate difference of the boundary point of the main grinding area, determine the real-time grinding width of the rail cross section at the current mileage coordinate.

[0020] Specifically, to overcome the mileage matching error caused by fluctuations in vehicle speed, a dual-probe differential logic combined with a dynamic spatiotemporal alignment algorithm is employed. Given the fixed physical mounting distance ΔL between the pre-scanning and rear-scanning laser profilometers on the vehicle body, and considering that the speed v of the grinding vehicle is not constant during grinding, a real-time speed feedback mechanism is introduced. The real-time travel speed v of the grinding vehicle is read in real time via a mileage encoder. w (t), the real-time operating speed specifically refers to the instantaneous travel speed of the grinding vehicle calculated based on the pulse signal from the odometer encoder and corrected for the center time timing. The system assigns the average speed within the sampling window to the center time of that window, rather than the end time, thereby eliminating the 0.5ΔT time lag caused by the traditional differential method and ensuring that the speed signal is strictly synchronized with the acquisition time of the laser sensor on the time axis. According to the formula Δt curr =△L / v w (t) Real-time calculation of time delay compensation. The system maintains a high-speed circular buffer to store the sequence of initial rail profile data before grinding at historical moments. For the real-time rail cross-sectional profile data G collected at time t... curr (t), the system dynamically backtracks to t'=t-△t curr At any given time, extract the corresponding initial reference profile data G of the rail. raw (t'). To ensure that the calculated geometric change features are strictly aligned with the original multi-source sensor data in the same spatial dimension, this step directly reuses the grid partitioning standard and index mapping mechanism established in the previous "Construction of Multi-Sensor Spatiotemporal Alignment Dataset". The system does not establish a new coordinate system, but treats the geometric features as a new data stream, based on the same window parameter (S). start , ΔS) and the same indexing rules, and place it into the existing structured grid cell.

[0021] During geometric feature calculation, the system, based on a time delay compensation mechanism, has established a pairing relationship between the real-time cutting data time t and the reference data time. At this point, the absolute mileage coordinate S corresponding to this paired data... t This refers to the same mileage reference value obtained through odometer data interpolation or mapping during the multi-sensor data preprocessing stage, as described earlier. In other words, S t The mileage coordinates S used in constructing the spatiotemporally aligned dataset mentioned earlier i Based on the same source, same reference, and spatiotemporal synchronization, no additional coordinate transformation is required. Based on the confirmed mileage coordinates S... t The system directly calls the grid index calculation logic defined earlier, that is, the rules used in the previous steps to determine the original data grid affiliation, combined with the starting mileage S of the current processing window. start Given the standard step size ΔS, S is analyzed. t The corresponding grid index k. Due to the input mileage coordinates S t Consistent with the original data, and using the same grid parameters (S) start Since ΔS and the window size are globally shared, the calculated index k is exactly the same as the grid index into which the original multi-sensor data falls at that moment. Boundary determination and feature aggregation are as described above. For valid features within the current window, the system encapsulates the geometric depth, width, and timestamp t into feature tuples and appends them to the geometric feature list of the k-th grid cell. The two sets of time-compensated data are spatially mapped to construct a mileage-paired dataset D. pair D pair (k)={(x k , z raw (k), z curr (k))|k=1,2,…,N}, where k represents the index of the transverse sampling point within the rail section, x k Let z be the lateral coordinate of the sampling point in the cross-sectional coordinate system. raw (k) represents the vertical coordinate of the initial reference profile at the sampling point, z curr (k) represents the vertical coordinate of the real-time profile at the sampling point, and N is the number of effective sampling points for a single frame profile, such as 1024 points. Regarding the horizontal coordinate x... k The source and specific construction process of spatial mapping are as follows: horizontal coordinate x kThe generation and unification of data: When the reference acquisition probe and the result acquisition probe acquire the rail profile, the initial output is raw measurement data (usually a combination of distance and angle) based on the sensor's own optical center. To eliminate the influence of probe installation position, tilt angle, and individual differences, the system first performs a unified coordinate transformation on the two sets of raw data. This process uses pre-calibrated probe installation parameters (including installation height, lateral offset, and tilt angle) to transform the local raw measurement values ​​of the sensor into a unified rectangular coordinate system of the standard rail section. In this standard coordinate system, the origin is defined as the top center of the theoretical centerline of the rail, and the horizontal axis extends along the width direction of the rail. After this transformation, the generated x... k This represents the absolute physical lateral position of the k-th sampling point relative to the theoretical center on the rail section. Because the two probes undergo rigorous synchronous calibration, for the same index k, their calculated lateral coordinates x... k Physically, it points to the same transverse position on the rail cross-section. The spatial mapping and dataset construction logic is not a complex image registration, but rather a data fusion process based on time synchronization and index alignment. After completing dynamic time delay compensation in the preceding steps, the system has locked two frames of data belonging to the same rail mileage position: one frame is the initial reference data frame collected at a historical time, and the other frame is the real-time data frame collected at the current time. Since the two frames of data have the same sampling frequency, the same field of view, and the same number of points N, and have both been converted to the aforementioned unified standard cross-section coordinate system, the system directly uses the index correspondence method for mapping: Positioning: The system simultaneously reads data points with the same index k from both the reference data frame and the real-time data frame; Extraction: Extracts the unified transverse coordinate x under this index. k Reference vertical height z raw (k) and real-time vertical height z curr (k); This combines the three values ​​into a triplet data unit. The system iterates through all sampling point indices from 1 to N, performing the above extraction and assembly operations sequentially, and finally arranges all triplets in index order to form a complete same-mileage paired dataset D. pair Based on the same mileage pairing dataset D pair The core features of the calculation are: 1) The real-time cutting depth h is calculated using the statistical quantile method, which is suitable for scenarios with high noise resistance requirements. It effectively filters out isolated noise points caused by dust and oil reflections, reflecting the true macroscopic cutting depth. The formula for calculating the real-time cutting depth h is: h = Percentile p (|z raw (k)-z curr(k)|), the preset quantile value p, preferably 95%, means that the depth difference of all sampling points is sorted by size, the top 5% of data points with the largest value are discarded, and then the maximum value is selected from the remaining 95% of data points as the final real-time cutting depth. 2) Calculation of real-time grinding width w. The grinding width represents the effective coverage of the grinding operation. The threshold connected component analysis method is adopted: a preset grinding threshold Th is set, preferably 0.1mm, which can be dynamically adjusted according to process requirements. Traverse all sampling points k and mark those that meet the condition (|z raw (k)-z curr Points where (k)|>Th are considered valid grinding points. On the horizontal x-axis, a connected component analysis algorithm is used to aggregate all adjacent valid grinding points into several independent grinding region sets. From these sets, the region with the largest horizontal span is selected as the main grinding region. Based on the main grinding region, the horizontal coordinate x of the leftmost valid point within this region is extracted. start And the rightmost valid point x end Calculate x start With x end The Euclidean distance between them represents the real-time grinding width. If multiple discontinuous grinding regions exist (e.g., double-sided overlapping grinding), the width of each region can be calculated separately or the union of the regions can be taken. The longitudinal direction is along the direction of the rail extension, which is the direction in which the grinding vehicle wheels roll (front-to-back direction). The transverse direction is perpendicular to the direction of the rail extension, i.e., from the left side to the right side of the rail, or from the inside to the outside. In this direction, the cross-sectional shape of the rail (e.g., railhead arc, rail web, rail base) unfolds. The sampling point index k is arranged along this transverse direction. The Z-axis, vertical, represents depth / height.

[0022] S200: Extract grinding drive features and multidimensional response features from each mileage grid point of the spatiotemporal aligned dataset, and calculate real-time grinding specific energy based on the mapping relationship between the two. If any of the multidimensional response features and real-time grinding specific energy of the current mileage grid point deviates from its corresponding expected response allowable range, then the grid point is marked as an anomaly. Based on the anomaly marking result, the data in the anomaly grid point is removed, and the data of the remaining mileage grid points are reorganized into a valid feature sequence. The multidimensional response features include at least the equivalent grinding force value, temperature rise value, and spectral energy value, and the grinding drive features include at least the statistical features of the geometric change features.

[0023] Specifically, during rail grinding operations, sensor data is inevitably contaminated by various non-grinding interferences. For example, oil or water film on the rail surface can lead to inflated laser profile measurements, while lubrication can cause underestimation of grinding force. Loose sensor connections under harsh conditions can cause instantaneous signal jumps or zeroing. Sparks and dust at the work site can also interfere with visual images. Unlike traditional data cleaning methods, this step is based on the principle of energy conservation in grinding thermodynamics, using grinding specific energy as a physical consistency criterion. Grinding specific energy is defined as the ratio of grinding power to material removal rate, and its physical meaning is the energy consumed to remove a unit volume of material. Under stable grinding conditions, regardless of changes in cutting depth, grinding width, or travel speed, grinding specific energy should remain within a reasonable range determined by material properties and grinding wheel characteristics. When non-grinding interferences such as oil, water film, or sensor disconnection occur, the physical coupling relationship between grinding force and cutting depth is disrupted, causing the calculated grinding specific energy to deviate significantly from this reasonable range. This step utilizes this physical law to mark grid points that deviate from the expected range as anomalies and remove them, only passing data that conforms to physical consistency to subsequent steps.

[0024] Furthermore, the sequence is recombined into a valid feature sequence, including: (1) For each standard mileage grid in the spatiotemporal aligned dataset, extract and calculate multidimensional response features from the multidimensional heterogeneous data associated with that grid point, wherein: The three-dimensional grinding force signal data of the grid point is statistically processed, and its mean value is defined as the equivalent grinding force value. The difference between the highest temperature value and the ambient reference temperature is extracted from the temperature field distribution image data and defined as the temperature rise value. The high-frequency vibration acceleration signal data is frequency domain transformed, and its power spectral density integral value is defined as the spectral energy value. Meanwhile, grinding drive features are extracted from the geometric change features associated with the grid point, including: calculating the average real-time cutting depth and the average real-time grinding width within the grid point. And obtain the real-time operating speed of the grinding equipment derived from the mileage encoding signal; (2) Based on the equivalent grinding force and the real-time running speed, calculate the instantaneous grinding power; based on the average real-time cutting depth, the average real-time grinding width, and the real-time running speed, calculate the instantaneous material removal rate; and based on the instantaneous grinding power and the instantaneous material removal rate, calculate the real-time grinding specific energy of the grid point. (3) Construct the measured state vector of the grid point, including the real-time grinding specific energy, temperature rise value, and spectral energy value, and perform two-level discrimination, wherein... First, determine whether the real-time grinding specific energy is within the preset reasonable range of grinding specific energy; Secondly, if it is within the reasonable range of the grinding specific energy, then according to the preset dynamic allowable range mapping relationship, find the allowable upper and lower limits corresponding to the real-time grinding specific energy, and determine whether the temperature rise value and the spectrum energy value exceed the range. (4) If any of the above judgment results are negative, the grid point and all corresponding data are marked as abnormal and the abnormal data is removed. If all the above judgment results are positive, the grid point and all corresponding data are marked as valid data. During the process of traversing all standard mileage grids, data marked as abnormal is removed in real time, and valid data is reorganized in order according to the arrangement of its corresponding standard mileage grids on the continuous actual mileage coordinate axis to form a valid feature sequence of multi-source synchronization.

[0025] Specifically, since the grinding machine travels at a certain speed v and the grid step size is ΔS, each discrete mileage grid point corresponds to a specific acquisition time segment in the time domain, i.e., a time window. The duration of this time window is approximately Δt = ΔS / v. The equivalent grinding force value F... eq The three-dimensional force signal includes the normal force F. n Tangential force F t and axial force F a Since grinding energy is mainly consumed in the tangential and normal directions, this embodiment uses a vector synthesis method to calculate the grinding force. And its mean value is defined as the equivalent grinding force. Subsequently, within the time window corresponding to the current grid point... The sequence is arithmetically averaged to obtain the equivalent grinding force value at each grid point, thus eliminating instantaneous fluctuations caused by high-frequency vibrations. For the temperature rise value, in each frame of the temperature field image time series, the highest temperature value across the entire image range is directly retrieved and recorded. This value represents the hottest state of the grinding contact area or spark generation at the current moment. From the same frame image, the average temperature of background areas far from the high-temperature grinding zone, such as image edges or unaffected rail surfaces, is selected, or the initial ambient temperature calibrated before operation is used as the environmental reference temperature T. env Subtracting the ambient reference temperature from the highest temperature value yields the net temperature rise after eliminating background interference. Spectral energy value E vib A Fast Fourier Transform (FFT) is performed on the high-frequency vibration acceleration signal to convert the time-domain signal to the frequency domain. The power spectral density is obtained by squared the amplitude in the frequency domain. Within a preset characteristic frequency band, such as 500Hz-5000Hz, which corresponds to the characteristics of abrasive grain breakage or chattering in grinding wheels, the power spectral density is integrated to calculate the spectral energy value. This integrated value can reflect the stability of the grinding process more sensitively than the simple time-domain peak value.

[0026] In this embodiment, the real-time grinding energy is calculated based on the physical principles of grinding, and the instantaneous grinding power is defined as the equivalent grinding force F. eq With real-time running speed v w The product of (t), i.e., P inst =F eq ×v w (t), this product represents the mechanical work consumed by the grinding system per unit time. Instantaneous material removal rate: defined as the volume of material removed per unit time, i.e., MRR. inst =a p ×b w ×v w (t), where a p b represents the average depth of cut in real time. w v represents the average real-time cutting width. w (t) represents the real-time operating speed of the grinding machine, v w (t) is the key physical quantity that converts the cutting section into volumetric flow rate. Similarly, if the velocity is zero, the material removal rate is zero. Real-time grinding specific energy E specific Grinding specific energy is defined as the ratio of power consumed per unit to material removal rate per unit, i.e.: Among them, in mathematical form, velocity v w It can be removed, but in the engineering implementation of this embodiment, the complete power and removal rate calculation process described above must be retained for the following reasons: P inst and MRR inst These represent two independent physical monitoring dimensions: energy input flow and matter output flow. Directly measuring or calculating these two intermediate variables helps to monitor anomalies in energy efficiency or removal efficiency separately in subsequent analyses. In the actual polishing process, F eq a p b w v w (t) are all dynamic quantities that fluctuate at high frequency with time. By introducing v w (t) participates in intermediate calculations, ensuring a strict match between the numerator (power) and denominator (removal rate) in the time domain and physical dimensions, avoiding errors caused by neglecting dynamic coupling effects due to direct static division. The grinding machine is in a stopped state (v w (t)≈0) or slow creeping v w (t) <v thresholdvw At that time, the material removal rate (MRR) inst When the value approaches zero, calculating the grinding specific energy will cause the value to diverge towards infinity or lose its physical meaning. Therefore, this embodiment includes a validity criterion: the system monitors v in real time. w If v w Below the preset effective work threshold v thresholdvwIf v < v, then skip the grinding specific energy calculation for the current moment, mark that moment as an ineffective grinding condition, and exclude it from the subsequent stability discrimination sequence. w ≥v thresholdvw Only when the time is right will the above division operation be performed to output a valid real-time grinding specific energy value.

[0027] In this embodiment, a high-frequency sampling time window ΔT is set. sample For example, if the timeframe is 10ms, count the number N pulses output by the encoder within that window. pulse Given that the encoder has K lines enc (Pulse / revolution), with the effective rolling radius of the wheel being R, calculate the theoretical displacement ΔL within this window. Calculate the average velocity within this window, v´=△L / ΔT sample Because the sampling window is extremely short, the change in vehicle acceleration within this tiny time period is negligible. To eliminate the half-beat time lag present in traditional calculation methods, this embodiment performs time alignment processing: the calculated average velocity v´ of the window is assigned to the center time t of the window on the time axis. center The sequence v generated is not the end of the window. w (t) represents the real-time walking speed after time alignment. w Each data point in the (t) sequence is strictly synchronized with the grinding current and laser profile data collected at the same time on the time axis. Considering that local unevenness on the rail surface, such as welds and scratches, may cause instantaneous micro-jumps or slight slippage of the measuring wheel, resulting in sudden changes in velocity pulses, the system performs secondary processing on the original velocity sequence: based on vehicle dynamics constraints, it calculates the instantaneous acceleration of the velocity sequence. If the absolute value of the acceleration at a certain moment exceeds the preset vehicle physical limit threshold, for example, >2m / s², the system will perform a secondary processing. 2 If the acceleration or deceleration capability far exceeds that of normal grinding operations, the point is determined to be a measurement outlier caused by wheel-rail impact, and is removed and repaired using interpolation of neighboring data. The sequence after outlier removal is then processed with median filtering or low-pass moving average to further filter out high-frequency noise.

[0028] In this embodiment, the physical stability criterion consists of two parts: the reasonable range of grinding specific energy and the mapping relationship between the dynamic allowable range. The reasonable range of grinding specific energy [E] min E max This range is determined based on the properties of the steel rail material being processed (such as hardness and toughness) and the characteristics of the currently used grinding wheel (such as grit size, bond, and hardness grade). Reference value E base This can be obtained through derivation based on grinding physics principles or through short-term on-site calibration experiments. Setting an allowable fluctuation tolerance Δδ, ±20%, to characterize reasonable fluctuations under normal operating conditions, the interval boundary is: E min =E base ×(1-δ), Emax =E base ×(1+δ), if the real-time grinding specific energy < E min If the grinding energy is greater than E, it is determined to be due to wheel slippage or air cutting (energy is not effectively used to remove material); max The grinding wheel is judged to be severely dulled, clogged, or experiencing intense friction (extremely low energy efficiency). Both of these situations are directly classified as unstable grinding states. The dynamic allowable range mapping relationship, in this embodiment, is within the reasonable grinding specific energy range [E...]. min E max Within this framework, upper and lower limits for the grinding specific energy and temperature rise were established. low (E), T high (E)] and the upper and lower limits of the spectral energy value [V] low (E), V high The one-to-one correspondence between [E] is achieved through a pre-set empirical mapping table or piecewise linear interpolation rules. The baseline data in the mapping relationship, namely the temperature rise and vibration thresholds corresponding to different grinding specific energies, are pre-set based on grinding process principles and historical experimental experience. During the initialization phase, technicians conduct multiple comparative experiments under standard operating conditions to statistically determine the temperature rise and vibration distribution ranges corresponding to normal grinding conditions at different grinding specific energy levels. These statistically obtained typical specific energy-threshold correspondence data pairs are directly and permanently stored as the system's basic lookup table or interpolation nodes. Alternatively, using a lookup table method, the system can assign a reasonable range of grinding specific energies [E] to the appropriate values. min E max The system is divided into several discrete gear levels, for example, each level is 5 J / mm³. When the real-time grinding specific energy falls into a certain level, the system directly calls the pre-stored threshold group (T) for that level. low T high V low V high This serves as the criterion. Alternatively, piecewise linear interpolation can be used. If the real-time grinding specific energy does not fall exactly at a preset level, the system automatically selects data from two adjacent preset levels and uses a conventional linear interpolation algorithm, i.e., calculating the median value according to the distance ratio, to smoothly calculate the unique threshold set corresponding to the current real-time grinding specific energy. This mapping relationship reflects the strong coupling characteristics of multiple physical quantities: in the low specific energy range, close to E... min This corresponds to relatively weak cutting action. At this point, the allowable upper limit for temperature rise and vibration is low, and the allowable lower limit approaches the ambient noise floor level. If the measured value exceeds this narrow band range, whether too high or too low, it is considered abnormal. High specific energy range, close to E... maxThis corresponds to a larger cutting load. In this case, the system automatically relaxes the upper limit of allowable temperature rise and vibration to accommodate normal heavy cutting thermal / vibration signals; simultaneously, based on the principle of energy conservation, a corresponding lower limit is set, meaning that heavy cutting is inevitably accompanied by a minimum temperature rise and vibration. If the measured value is lower than this lower limit, it indicates sensor failure or data synchronization anomaly. For each mileage grid point, two lines of defense checks are performed according to the following priority: First line of defense: Energy efficiency check: Calculate real-time grinding specific energy E. specific If the judgment is If the grid point is deemed invalid, subsequent checks are skipped. Enter the second line of defense, and based on the current E specific The value query mapping relationship is used to obtain the corresponding dynamic threshold group. The second line of defense: multi-source consistency check, temperature rise consistency: check whether the measured temperature rise ΔT meets the T... low ≤ΔT≤T high If ΔT>T high This is determined to be an overheating anomaly. If ΔT <T low The following conditions are considered abnormal: a sensor malfunction or a physical logic contradiction, such as no temperature rise during heavy cutting. If any condition is not met, the system is deemed invalid. Vibration consistency: Check the measured spectral energy V. real Does it satisfy V? low ≤V real ≤V high If V real >V high The condition is determined to be a severe vibration anomaly. If V real <V lowThe result is determined to be either a vibration sensor failure or data loss. If either condition is not met, the result is deemed invalid. Valid feature sequence output: Grid points that pass both of the above checks are marked as valid, and their original data and feature vectors are retained and entered into the subsequent reconstruction and calculation process. Invalid data removal: Any grid point that fails the check is marked as invalid, and the data within that grid point is directly removed in subsequent data processing to ensure the physical reliability of the final result. This embodiment uses a short-term adaptive calibration method in the field: Calibration condition selection: Before the formal grinding operation starts, the control system automatically selects a section of rail with a relatively uniform surface condition and no obvious defects as the calibration area. Representative process parameters are set, including standard cutting depth and feed rate, and the grinding actuator is controlled to perform N repetitive stable grinding tests, such as 3-5 times. The grinding force, actual cutting depth, and grinding width of each test are collected in real time. The grinding specific energy of a single test is calculated, and outliers in the grinding specific energy sequence that deviate from the mean by more than a preset threshold, such as 3 times the standard deviation, are removed to eliminate the influence of transient or accidental interference at startup. The arithmetic mean or median of the remaining data after cleaning is determined as the benchmark grinding specific energy under the current working conditions, and an allowable fluctuation range is constructed based on the engineering experience tolerance coefficient δ, δ∈[15%,25%]. This method does not rely on cloud-based historical databases or offline large-scale test regression. It only needs to use a very short period of field measurement data at the beginning of the operation to generate personalized criteria specific to the current grinding wheel-rail-environment system. Even if the grinding wheel model or rail material batch is changed, the system can quickly adapt by re-executing this short-term calibration program, which has strong field adaptability and deployment flexibility. Dynamic reorganization refers to constructing a new, logically continuous, and effective feature sequence. During reorganization, the original mileage coordinates of each data point are retained as attributes, but the data is compactly arranged in memory or arrays. During the reorganization process, the system removes data from grid points marked as invalid, but retains the original mileage coordinates of the remaining valid grid points as attribute information. Subsequently, the valid grid points carrying the original mileage coordinate attributes are compactly arranged in storage space, forming a valid feature sequence with a continuous index but dynamically changing physical mileage intervals. When calculating the proportion of abnormal grid points, the system first reads the original mileage coordinates of the first and last valid grid points within the mileage processing unit corresponding to the current iteration cycle, and calculates the physical span between them. Then, based on the physical span and a preset standard grid resolution, the system estimates the theoretical number of grid points that should theoretically be included within that span. Finally, the system calculates the difference between the theoretical number of grid points and the actual number of valid grid points within the unit, and defines the ratio of this difference to the theoretical number of grid points as the proportion of abnormal grid points in that mileage processing unit.

[0029] S300: Construct an iterative processing cycle, obtain the adaptive reference value of each feature branch in the previous cycle, determine the deviation feature of the feature value of each feature branch in the current cycle relative to the adaptive reference value based on the adaptive reference value, take the grinding specific energy distribution feature in the current cycle as the global prior constraint, combine the deviation feature, dynamically adjust the weight of each feature branch in the current cycle, and characterize the result of the weighted fusion of the feature values ​​of each feature branch as the comprehensive state value of the current cycle. The iterative processing cycle is formed by dividing the effective feature sequence into several continuous processing units based on the data distribution characteristics of the effective feature sequence. The feature value of each feature branch characterizes the transient thermal shock degree, mechanical vibration instability trend and abnormal surface morphology state in the current cycle.

[0030] Specifically, step S300 is the core step in this embodiment to achieve intelligent perception and adaptive evaluation of the grinding state. It adopts a dual adjustment mechanism: an individual enhancement strategy, which independently amplifies the fusion weight of a single branch feature for sudden deviations. This allows the system to still capture early defects in a single dimension when other indicators are normal. The global gain strategy adjusts the weight base of all branches synchronously according to the overall energy load state. Under high load conditions, the system appropriately reduces sensitivity to prevent noise; under low load conditions, the system increases sensitivity to detect minor faults.

[0031] Furthermore, the weighted fusion of the feature values ​​of each feature branch is represented as the comprehensive state value of that period, including: The iterative processing cycle is based on the effective feature sequence, which is divided into several consecutive mileage processing units according to a preset sequence step size, and the data processing process of each processing unit is mapped to a time iteration cycle of the control system; wherein, the processing unit is composed of consecutively arranged effective grid points in the effective feature sequence; Iterate through the given iteration cycles, and for the current iteration cycle, execute the following steps in real time: (1) Read the original mileage coordinates of the first and last effective grid points in the unit and calculate the physical span. Based on the physical span and the preset standard grid resolution, calculate the number of theoretical grid points that should be included in the span. Calculate the difference between the number of theoretical grid points and the actual number of effective grid points in the mileage processing unit, and determine the ratio of the difference to the number of theoretical grid points as the proportion of abnormal grid points. Based on the mileage processing unit corresponding to the current iteration processing cycle, extract the real-time grinding specific energy values ​​of all mileage grid points in the unit, and calculate the arithmetic mean of the values ​​as the representative grinding specific energy in the cycle. Based on the temperature field distribution image corresponding to each mileage grid point in the mileage processing unit, the ratio of the temperature rise difference between adjacent grid points to the mileage distance is calculated to obtain the temperature rise gradient sequence. The maximum value in the temperature rise gradient sequence is taken as the transient thermal shock branch characteristic value in this period. Extract the spectral energy values ​​of all mileage grid points within the mileage processing unit, and calculate the root mean square value of the spectral energy values ​​as the characteristic value of the mechanical vibration instability branch within that period; Color space conversion and threshold segmentation are performed on the visual image of the polishing area of ​​each mileage grid point in the mileage processing unit. The proportion of single-point thermal damage area of ​​the grid point is calculated. The proportion of single-point thermal damage area of ​​all mileage grid points in the current cycle is compared. The maximum value is selected as the thermal damage morphology anomaly degree of the current cycle. The thermal damage morphology anomaly degree is used as the surface morphology anomaly branch feature value of the current cycle. (2) Construct a multi-dimensional state observation set, execute a hierarchical weighted fusion strategy, and obtain a comprehensive state value generated by multi-source feature fusion.

[0032] Specifically, to address the data processing challenges caused by fluctuations in the grinding machine's operating speed and instability in local grinding conditions, this invention employs a dynamic iterative processing mechanism based on effective feature sequences. The core of this mechanism lies in defining a complete computational closed loop of the control system—a time iteration cycle—based on the continuity of the effective grinding state. Since data within grid points marked as ineffective have been removed, the remaining effective data is logically rearranged to form a multi-source synchronous effective feature sequence. At this point, adjacent points in the sequence may be discontinuous in terms of physical mileage, resulting in mileage gaps. To balance computational real-time performance and data quality, the system divides the aforementioned effective feature sequence into several continuous processing units.

[0033] In a preferred embodiment, a processing unit is defined as containing a preset number N valid grid points. Window advancement mechanism: the starting point of the k-th processing unit is immediately followed by the ending point of the (k-1)-th processing unit, i.e., the next valid point in the sequence. The value of N is determined comprehensively based on the surface precision requirements of the polishing process, the sampling frequency of sensor data, and the computing power of the vehicle controller. For example, when N is larger, the calculated feature statistics are more stable, but a certain computational lag is introduced; when N is smaller, the response is more sensitive, but the noise resistance is relatively weaker. In this embodiment, N is preferably set to an integer between 10 and 50 to achieve a balance between real-time performance and stability. Due to the removal of abnormal data from the sequence, the N valid grid points within a mileage processing unit may be discontinuous in physical space. The system calculates the actual physical span covered by the unit by reading the original mileage coordinate attributes of the first and last valid grid points within the unit. If poor grinding quality in a certain section leads to the rejection of a large amount of data, the difference between the first and last coordinates of the processing unit, i.e., the actual physical span, will automatically increase in order to gather N valid points. This logically skips the physical intervals corresponding to invalid data. Skipping abnormal intervals: When dividing the window, the system automatically skips the physical mileage intervals corresponding to the rejected abnormal grid points and does not include them in the data volume of the processing unit.

[0034] The computation timing of the control system is triggered by the cumulative progress of the effective sequence. Starting from the system's acquisition of the first effective point required by the k-th processing unit, the physical time process until the termination condition of the processing unit is met, such as collecting N effective points and calculating the result, is defined as the k-th time iteration cycle of the control system.

[0035] In this embodiment, within the current iteration cycle, the system performs the following aggregation operation on the data of all grid points: First, it directly calls the real-time grinding specific energy value calculated in the previous step. The system traverses all grid points within this cycle, extracts the real-time grinding specific energy sequence, and calculates its arithmetic mean. This average value is defined as the representative grinding specific energy of the current cycle. This value, by smoothing single-point noise, macroscopically characterizes the overall energy load level of the current contact interface between the grinding wheel and the rail, serving as a global prior basis for subsequent weight allocation. Second, the difference in temperature rise between two adjacent grid points in the sequence is calculated; then, the ratio of this temperature rise difference to the mileage distance between adjacent grid points is calculated to obtain the local spatial temperature gradient between these adjacent point pairs. Since the grinding machine is operating, the spatial temperature distribution changes in relation to the temporal temperature rise rate; therefore, the local spatial temperature gradient can characterize the rapid temperature rise phenomenon on the rail surface within a very short spatial and temporal range, i.e., transient thermal shock. After completing the calculation of the local spatial temperature gradient for all adjacent point pairs within the window, the system iterates through the obtained gradient value set and selects the maximum value as the transient thermal shock branch characteristic value for the current period. Third, the system first calls the spectral energy value sequence of each grid point within the current window, which has been solved in the previous steps, where each spectral energy corresponds to the power spectral density integral value of the grid point's vibration signal. Subsequently, the system performs a root mean square aggregation operation on the energy concentration of the sequence: squaring each spectral energy value in the sequence, calculating the arithmetic mean of all squared values, and then taking the square root of the mean. The calculated root mean square is defined as the characteristic value of the mechanical vibration instability branch in the current cycle. Fourth, the system directly processes the original visual data in real time within the current iteration cycle. First, it acquires the original visual image sequence of the grinding area corresponding to all grid points within the current time window. For each frame of the image in the sequence, the system performs the following image processing steps: converting the original RGB color space image to the HSV (hue-saturation-luminance) color space, which is more sensitive to color, to eliminate the interference of ambient light changes on burn color recognition; setting a threshold for binarization segmentation based on the characteristic hue range of rail thermal damage discoloration (such as blue or purple oxide layers), and extracting the mask of the suspected thermal damage area; counting the total number of pixels in the suspected thermal damage area in each grid point image, and calculating its proportion to the total number of pixels in the effective grinding area of ​​that grid point, to obtain the proportion of thermal damage discoloration area at that point. After processing all grid points within the window as described above, the system obtains a sequence of thermal damage discoloration area proportions. Subsequently, the system performs an extreme value retrieval operation for the worst local state on this sequence: iterating through all values ​​in the sequence, directly selecting the maximum value, and defining this maximum value as the surface morphology anomaly branch feature value for the current period, i.e., the thermal damage morphology anomaly degree. In rail grinding quality assessment, severe local burns are often more destructive than the overall average state. By directly performing the calculation from image to proportion in the current step and selecting the maximum value instead of the average, the system avoids the dilution effect of large-area normal region data on severe local defects, ensuring zero-tolerance capture of surface integrity risks and preventing missed detections. Through the above processing, the system constructs a multi-dimensional state observation vector containing representative grinding specific energy, transient thermal shock branch feature values, mechanical vibration instability branch feature values, and surface morphology anomaly branch feature values.

[0036] This embodiment provides a dynamic reference value initialization and convergence method based on physical limit mapping. By mapping the physical range of each sensor to the theoretical limit value of its characteristic parameters, a highly fault-tolerant initial reference value is set, effectively achieving anti-interference in the startup phase and high-sensitivity detection in the operation phase. First, for each monitoring branch, a mapping relationship between the physical range and its corresponding theoretical limit value of the characteristic value is established. For the mechanical vibration instability branch, its characteristic value is the root mean square value of the spectral energy. The system regards the physical safety fuse line of the vibration sensor, such as 90g, as the time-domain limit amplitude. Based on signal processing principles, this time-domain limit is mapped to the theoretical limit spectral energy value in the frequency domain. This mapping relationship represents the theoretical upper limit of the spectral energy corresponding to when the vibration intensity reaches the physical red line. For the surface morphology anomaly branch, its characteristic value is the thermal damage morphology anomaly degree, that is, the proportion of single-point thermal damage area, with a value range of 0~100%. The system defines the physical limit of a drastic phase transition in the H channel of the image sensor, such as 160°, as the anomaly detection threshold in feature calculation. When all pixels in the entire field of view exceed this threshold, the corresponding theoretical limit of the feature is 100%, meaning a hue jump occurs across the entire field of view. For the transient thermal shock branch, its characteristic value is the maximum value of the temperature rise gradient sequence. The theoretical upper limit of this characteristic value, i.e., the theoretical maximum temperature rise gradient value, is derived based on the physical saturation value of the thermal infrared sensor, i.e., the upper limit of its range and its spatial resolution or time response frequency. Specifically, the theoretical maximum temperature rise gradient value can be defined as the limiting rate of temperature change that the sensor can capture within one sampling period or one pixel unit, and its value cannot exceed the ratio of the sensor's upper limit of range to its resolution. Secondly, based on the above theoretical limits, the initial reference value for the first iteration period is set. To effectively filter sensor noise and tolerate transient changes in operating conditions during the initial startup phase, this embodiment employs a high-tolerance range strategy. The initial reference values ​​are set to 40%–60% of the theoretical limits. Specifically, the initial reference value for the mechanical vibration branch is set to 0.4–0.6 × the theoretical limit spectral energy value; the initial reference value for the surface morphology anomaly branch is 0.4–0.6 × 100%, allowing for area fluctuations of 40%–60%; and the initial reference value for the transient thermal shock branch is set to 0.4–0.6 × the theoretical maximum temperature rise gradient value. This strategy creates an extremely lenient safety buffer, ensuring that even if significant energy fluctuations, changes in illumination, or temperature gradients occur during system startup, as long as the physical limits are not exceeded, the startup process is considered acceptable, thus fundamentally avoiding false alarms. Finally, after completing the first monitoring cycle, the system immediately initiates a dynamic convergence mechanism based on residual analysis.

[0037] Furthermore, a layered weighted fusion strategy is implemented, including: Based on the current iteration processing period T k Representative grinding specific energy within The linear normalization method is used to map them to global prior weight coefficients. The calculation formula is as follows: ; Where (k) is the index of the current iteration cycle, (k) = 1, 2, ..., E min and E max These are the preset lower and upper threshold values ​​for grinding specific energy under normal grinding conditions, respectively, P min and P max These are the preset lower and upper limits for the global prior weight coefficients, respectively. `clip(·)` represents the truncation function, used to restrict the calculation results to [P]. min P max Within the interval; Get the current iteration processing period T k The eigenvalues ​​of each feature branch are read from memory, and the previous iteration period T is read from memory. k-1 Updated and stored baseline values ​​for each feature branch Calculate the normalized deviation The calculation formula is as follows: ; Where i=1,2,3, i=1 represents the transient thermal shock branch, i=2 represents the mechanical vibration instability branch, and i=3 represents the morphological anomaly branch. Based on the current iteration processing period T k Global prior weight coefficients and the normalized deviation Calculate the adaptive fusion weights of the i-th branch. The calculation formula is as follows: ; Where α is the deviation response sensitivity coefficient, β is the minimum basic weight coefficient, γ is the gain coefficient, n is a nonlinear sensitivity adjustment factor greater than 1, and K th K is the preset overall energy load warning threshold. th ∈[E min E max ]; In the current iteration processing period T k Internally, the calculated weights of each branch are used for adaptive fusion. The eigenvalues ​​of the transient thermal shock branch, the mechanical vibration instability branch, and the abnormal surface morphology branch are weighted and fused to obtain a comprehensive state value. This comprehensive state value characterizes the overall load level determined by grinding specific energy, thermal shock, mechanical vibration, and surface morphology under the current iterative processing cycle.

[0038] Specifically, firstly, the system is based on the current iteration period T. k Representative grinding specific energy collected internally Calculate the global prior weight coefficients The system utilizes a preset energy range for normal grinding conditions [E]. min E max A linear normalization mapping is performed on the real-time energy. During the mapping process, a preset lower limit value P of the global prior weight coefficients is introduced. min With upper limit value P max Perform linear interpolation. In this embodiment, P min Physical definition: Noise floor shielding threshold logic: P min The purpose is not to limit the weights, but to shield the sensor's background noise. Even when the grinding specific energy is under absolute no-load or extremely light-load conditions, the sensor still has background noise. If the weights approach zero, it will cause instability in the denominator of the fusion formula or the loss of weak signals. Therefore, P min This is set as the critical value that ensures the system maintains the minimum observation gain even in a purely noisy environment. It is typically taken as the weighting ratio corresponding to the inverse of the sensor signal-to-noise ratio, or the minimum stable gain in engineering control theory, ranging from 0.05 to 0.1, to ensure the system never goes blind. max Physical definition: Saturation protection and overexcitation prevention threshold, P max This is to prevent system crashes or overreactions caused by single-point failures. Even with extremely high grinding energy, approaching the physical melting point, it's undesirable to infinitely amplify the weight of any particular branch, lest a momentary spike from a single sensor distort the entire fusion result. Therefore, P max This is set as a safety redundancy factor to withstand transient interference while ensuring high sensitivity. In this embodiment, P is preferably set... min =0.1, P max =0.9, P min =0.1 Even under no-load or micro-cutting conditions, the system retains a 10% base observation gain to prevent numerical instability caused by the denominator approaching zero and to shield against false triggering caused by sensor noise floor; while setting P max =0.9, even under extreme full-load conditions, forcibly retaining at least 10% of the discourse power of other feature branches. Finally, the calculation result is strictly limited to the interval [0.1, 0.9] by the truncation function clip(·), ensuring... It exhibits a strict positive correlation with the overall energy load and changes smoothly.

[0039] Secondly, for the three characteristic branches of transient thermal shock (i=1), mechanical vibration instability (i=2), and abnormal surface morphology (i=3), the system obtains the current period T respectively. k Statistical aggregate value of feature values ​​of each grid point within the grid To accommodate feature drift caused by grinding wheel wear and batch variations in workpiece materials, this embodiment does not use a fixed threshold, but instead calls the threshold value from the previous iteration cycle T. k-1 Updated dynamic baseline value To calculate the normalized deviation This deviation reflects the relative rate of change of the current feature with respect to the recent health status center. When the feature value deviates from the baseline value, the deviation increases linearly with the degree of deviation; the greater the deviation, the higher the fusion weight of that branch.

[0040] Subsequently, the system is based on the aforementioned global prior weight coefficients. and normalized deviation The adaptive fusion weights of the i-th branch are calculated through a dual dynamic adjustment mechanism. The calculation of this weight involves two coupled terms: the first term is the basic response term. The key lies in the selection of the nonlinear sensitivity adjustment factor n, which is used to construct the isolation band between normal fluctuations and real faults. When n is close to 1, the system is too sensitive to small noises and prone to false alarms; when n is too large (e.g., >10), the system is slow to react to early minor faults and prone to missed alarms; the optimal range is usually n∈[2,4]. Specifically, for the transient thermal shock branch, considering that the thermal accumulation effect of grinding burns conforms to the square law of energy, n=2 is preferred; for the mechanical vibration instability branch, considering the cubic relationship between crack propagation and stress concentration, n=3 is preferred; for the surface morphology anomaly branch, considering that visual feature mutations often have step characteristics, n=2 is preferred. This nonlinear setting of n>1 gives its weighting function a soft threshold characteristic. When the deviation is small, The item decayed sharply, and the weight was mainly determined by This decision effectively suppresses noise interference; and when The term increases non-linearly, and the greater the deviation, the faster the growth rate, quickly highlighting fault characteristics. The second term is the global gain term. , The warning threshold is based on the process safety margin, according to the formula. Confirmed. This threshold is calculated based on the current process parameters (rail material, grinding wheel type, target cutting depth) during the task initialization phase and remains fixed throughout the same batch of operations. λ is the safety warning lead time coefficient, preferably ranging from 0.6 to 0.8. For example, if λ = 0.7, when the real-time grinding ratio reaches 70% of the normal operating range, the system determines that it has entered a high-sensitivity monitoring state. At this time, if Exceed The global gain term will be greater than 1, acting as a common multiplier to synchronously amplify the adaptive fusion weights of all three branches. High energy loads are often a common precursor to multiple failure modes. Simultaneously increasing the sensitivity of all sensors in high-risk areas can enable early interception of coupled faults.

[0041] The minimum basic weight coefficient β is used to set the basic weight of the system under fault-free and low-load conditions, ensuring that the monitoring channel remains active under normal operating conditions and avoiding the formation of sensing blind spots. Its preferred value range is 0.01 to 0.15. If β is too small, the monitoring channel will tend to shut down under normal operating conditions; if it is too large, background noise will be excessively amplified. In this embodiment, the typical value of β is 0.05, meaning that even in the most stable grinding stage, the system still retains 5% of the basic weight to maintain the baseline response of the monitoring channel. Secondly, the deviation response sensitivity coefficient α determines the growth rate of the weight as the degree of deviation from the fault characteristics increases. To ensure that the system can respond quickly when an anomaly is detected, and to avoid the weight saturating under slight disturbances due to excessive sensitivity, this embodiment applies a linkage constraint to α and β: requiring the total value of the basic response term when the normalized deviation is 1 (i.e., β + α) to be controlled between 0.5 and 1.0. Based on this logic, when β is 0.05, the preferred value range of α is limited to between 0.45 and 0.95. In this embodiment, the typical value is 0.9. Under moderate deviation, the basic response term is approximately 0.95, which is significantly higher than the background noise while providing ample dynamic space for subsequent nonlinear growth. Finally, the global gain intensity coefficient γ is used to control the synchronous amplification factor of all branch weights in the high-load warning region. Its value is not set independently but is based on a preset value. A reverse calculation is performed to ensure that the total amplification factor from the global gain term is strictly controlled between 1.5 and 2.5 times when the real-time energy load reaches the process limit. For example, when When the value is 0.7, to ensure the maximum amplification factor under full load is approximately 1.9 times, the value of γ should satisfy 1 + γ × (1 - 0.7) ≈ 1.9, which yields γ approximately 3.0. Therefore, under this configuration, the preferred value range for γ is 1.7 to 5.0. In this embodiment, γ = 3.0 is ultimately selected. This allows the system to automatically increase the weight of all fault characteristics by approximately 1.9 times when entering high-risk operating conditions, thereby increasing alertness while avoiding excessive system overreaction caused by excessive gain. Through the limitation of the above parameter range and the construction of interrelated logic, this scheme enables the adaptive fusion weight to remain stable in the low value range under normal operating conditions to filter out noise, smoothly jump to the medium-high value to highlight characteristics when a single fault occurs, and moderately amplify under high-load extreme operating conditions to achieve the highest level of early warning, thus constructing a hierarchical, dynamically reasonable, and numerically stable adaptive monitoring system. The adaptive fusion weight in this embodiment... This represents the gain amplification factor of the characteristic signal, not the probability distribution value. Therefore, the sum of the weights of each branch does not need to be normalized to 1. Under normal operating conditions, the sum of the weights is much less than 1 to suppress noise; under fault or high-load conditions, the sum of the weights can be much greater than 1 to significantly highlight abnormal features.

[0042] Finally, in the current iteration period T k Internally, the calculated weights of each branch are used for adaptive fusion. The system performs weighted summation of feature data from three branches—transient thermal shock, mechanical vibration instability, and abnormal surface morphology—to generate a comprehensive state vector characterizing the evolution trend of microscopic damage. This vector not only integrates multi-dimensional sensor information but also embeds the current load level and health baseline context, achieving high-precision, low-false-alarm monitoring of grinding conditions without relying on a large amount of historical training data models.

[0043] S400: Based on the comprehensive state value and its historical evolution trend of the current iteration processing cycle, an environmental noise observer is constructed to calculate the noise benchmark that characterizes the degree of environmental interference in real time. The noise benchmark is used to adaptively correct the benchmark values ​​of each branch feature value of the previous iteration processing cycle to obtain the adaptive benchmark values ​​of each feature branch of the current iteration processing cycle, which are used for the next iteration processing cycle.

[0044] Furthermore, the adaptive baseline values ​​for each feature branch in the current iteration processing cycle are obtained, including: The current iteration processing period T is generated using a first-order exponential smoothing algorithm. k Predicted comprehensive state value The formula is as follows: ; in, For smoothing coefficients, The previous iteration processing period T k-1 The overall state value, Previous iteration processing cycle T k-1 The predicted comprehensive state value; Using the current iteration processing period T k Comprehensive state value Combined with predicted state values Calculate the current iteration processing period T k absolute value of prediction residuals The formula is as follows: ; Calculate the moving average of the absolute values ​​of the prediction residuals over the most recent M periods, and use it as the current iteration period T. k Dynamic noise benchmark The formula is as follows: ; Where M is the preset sliding window length, M>3, and j is the temporal index variable within the sliding window, representing the j-th cycle backtracking from the current iteration processing cycle; Calculate the current iteration processing period T k The ratio of the absolute value of the predicted residual to the dynamic noise baseline yields the relative disturbance amplitude. ,in, To prevent division by zero, the value is set to... The formula is as follows: ; Construct global scaling factor ,in, The sensitivity coefficient is given by the following formula: ; Using the global scaling factor For the previous iteration period T of each of the three feature branches k-1 benchmark value Perform synchronous updates to obtain the baseline values ​​of each feature branch updated in the current iteration cycle. The formula is as follows: ; The current iteration processing cycle benchmark value Write to memory as the next iteration period T k+1 The normalized deviation criterion is as follows: if the current period is the initial period, then the preset initial criterion value is used as the adaptive criterion value for the current period.

[0045] Specifically, this embodiment uses a comprehensive state value. This allows for the perception of the current polishing environment and the dynamic adjustment of fault detection thresholds for each branch. First, a first-order exponential smoothing algorithm is used to generate the current iteration period T. k The predicted comprehensive state value is calculated. Here, the smoothing coefficient is set to 0.5, which is determined based on the frequency domain characteristics of the rail grinding condition. Effective fault symptoms, such as wheel passivation and continuous burning, usually manifest as low-frequency trend changes, while environmental disturbances, such as ballast impact and sensor noise, manifest as high-frequency transient spikes. A smoothing coefficient of 0.5 can achieve a balance between preserving trend information and suppressing high-frequency noise, so that the predicted comprehensive state value can stably reflect the theoretical steady state of the system under undisturbed conditions. Furthermore, the absolute value of the prediction residual is calculated, which quantifies the degree to which the actual state of the system deviates from its theoretical steady state at the current moment. In order to eliminate the influence of dimensions and assess the severity of the current disturbance, this embodiment calculates the relative disturbance amplitude, which is essentially a disturbance intensity index. When this occurs, it indicates that the current fluctuation falls within the normal background noise range. This indicates that the system is experiencing an abnormally severe transient shock, such as when a grinding machine passes over a rail joint or turnout, with fluctuations far exceeding normal levels. This indicates that the system is in an extremely stable operating condition, with current fluctuations far below the background noise level. A global scaling factor is constructed using the hyperbolic tangent function tanh to achieve nonlinear adjustment of the reference value. It can map drastic changes in the input variable to a finite, smooth output range. This ensures that the adjustment of the reference value is gradual and controlled, avoiding system oscillations caused by sudden parameter changes. Zero-point drift of R(k)-1: Since R(k)=1 represents the normal noise level, the input is subtracted by 1 so that when the system is in a normal fluctuation state, tanh(0)=0, at which point λ(k)≈1, that is, the reference value remains unchanged. Sensitivity coefficient σ: The value is set to 1.0 to adjust the steepness of the function curve, ensuring that the scaling factor can linearly respond to residual changes within the typical disturbance intensity range. The value range is [0.8, 2.0]. When the system detects a severe disturbance, i.e., R(k) is very large, the grinding state itself will fluctuate drastically. At this time, if the original strict discrimination criterion is maintained, it is very easy to misjudge this normal operating condition fluctuation as equipment failure. Therefore, a positive adaptive adjustment mechanism is designed: when R(k) increases, the tanh term aligns to 1, and the driving coefficient... The tendency towards the upper limit of 2.0 indicates that the system amplifies the discrimination benchmark to twice its original value. When R(k) is extremely small, the tanh term aligns with -1, and the driving coefficient... It tends towards the lower limit of 0.8. Finally, using the global scaling factor... Update the baseline values ​​for each feature branch.

[0046] S500: Within the current iteration processing cycle, based on the abnormal grid point ratio, comprehensive state value, and evolution characteristics of the adaptive baseline value determined by the abnormal marking results, multi-dimensional collaborative verification is performed; based on the working condition confidence level generated by the verification results, a hierarchical response strategy is triggered to achieve closed-loop intelligent monitoring of grinding operations.

[0047] Furthermore, hierarchical control instructions are generated to adjust the grinding operation status, including: Within the current iteration cycle, key indicators are extracted, including: the proportion of abnormal grid points, the overall status value, and the trend of dynamic benchmark value, among which the trend of dynamic benchmark value. Maintain a historical data queue for the most recent N periods, where N ≥ 10, and the queue contains the key indicators. Perform the following adaptive validation: (1) Data quality verification: The proportion of abnormal grid points in the current iteration period is sorted and compared with all values ​​in the historical queue. If the abnormal grid point in the current iteration period ranks first in the historical queue and the difference between it and the second historical grid point exceeds the preset significance ratio, it is determined that the data quality is seriously abnormal. (2) Comprehensive risk verification: The comprehensive status value of the current iteration period is sorted and compared with all values ​​in the historical queue. If the comprehensive status value of the current iteration period ranks among the top two in the historical queue, it is determined that the comprehensive risk has increased. If the comprehensive status value of the current iteration period ranks first in the historical queue and remains first for two consecutive periods, it is determined that the comprehensive risk is serious. (3) Trend anomaly verification: Calculate the number of cycles in which the dynamic benchmark value changes in the same direction. If the dynamic benchmark value changes in the same direction for L consecutive cycles, L≥5, and the current dynamic benchmark value changes in the opposite direction to the historical main trend, it is judged as a benchmark trend reversal warning. Based on the above verification results, the following hierarchical response strategy will be implemented: Level 1 warning: When any verification item is determined to be abnormal or elevated, a warning signal is generated; Level 2 adjustment: When any verification item is determined to be serious, or when both data quality verification and comprehensive risk verification are determined to be abnormal / increased, an automatic intervention instruction is generated; Level 3 Emergency Stop: When both data quality verification and comprehensive risk verification are deemed critical, or when the baseline trend reversal warning and any critical determination occur simultaneously, an emergency stop command is generated. The generated hierarchical control commands are sent to the grinding equipment actuator to adjust the grinding operation status. At the same time, abnormal events of the current cycle are recorded in the log. If a level 3 emergency stop is triggered, the system automatically saves a number of complete data snapshots before and after the fault and reports the key information of the abnormal event to the cloud operation and maintenance platform.

[0048] Specifically, the core of this embodiment is to construct a multi-dimensional, hierarchical dynamic verification and response mechanism. In the current iteration cycle T... k Internally, the system acquires three key indicators: the percentage of abnormal grid points. Comprehensive state value and the changing trend of dynamic benchmark values , Regarding the percentage of abnormal grid points This indicator reflects the degree of data loss or anomalies within the current grinding section. A high percentage indicates extremely unstable grinding conditions or severe sensor interference in that section. Data from the most recent N periods constitutes a profile of the equipment's normal behavior under current operating conditions. By maintaining this sliding window, the system can automatically adapt to slow shifts in operating conditions, such as slow tool wear, without false alarms due to minor deviations in the baseline value. Subsequent relative sorting verification requires a sample space. Without a historical queue, it's impossible to determine whether the current value is a normal high or an abnormally high value.

[0049] First, data quality verification is crucial. During grinding operations, sensor data is highly susceptible to interference from coolant, metal dust, and electromagnetic interference. Directly using contaminated data for status assessment will lead to erroneous decisions. Therefore, the reliability of the data must be quantitatively assessed first. This embodiment employs a dual judgment mechanism: a pre-set ratio and a base threshold. The pre-set ratio, for example, the top 10%, is based on quantile statistics. Due to differences in workpiece material and hardness, background noise levels are dynamically changing. By setting a relative ratio, such as the top 10%, the system constructs an adaptive dynamic threshold. If the current percentage of abnormal grid points is within the worst 10% range relative to historical performance, it is considered abnormal. This solves the problem of poor adaptability of fixed thresholds under different operating conditions. The base threshold, such as a 30% setting, is based on the Nyquist sampling theorem and engineering experience. When invalid points exceed 30%, the effective sampling rate can no longer support the confidence level of feature extraction. This is to prevent security vulnerabilities caused by the failure of the relative sorting mechanism under extreme conditions, such as complete sensor failure. For severe anomaly detection, the design prioritizes the highest percentage with a significant difference. This design is based on outlier detection of step-like collapses, such as sudden sensor detachment. This design effectively distinguishes between slow deterioration and sudden failures, improving response timeliness. A significant difference means that the relative difference between the percentage of outlier grid points in the current iteration cycle and the percentage of the second-highest percentage of outlier grid points in the historical queue must be greater than or equal to 50% of the percentage of the second-highest percentage of outlier grid points in the historical queue. During grinding, the increase in outliers caused by normal wear of the grinding wheel and minor fluctuations in the workpiece material usually exhibits a linear or exponential gradual characteristic, with an increase of less than 20% per cycle. However, when catastrophic failures occur, such as sensor detachment or lens being completely blocked by coolant, the percentage of outliers can jump instantly from a low level, such as 5%, to a high level, such as 90%. Setting a 50% significance threshold is precisely the watershed between the gradual and abrupt change zones, accurately capturing such step-like collapses while ignoring normal fluctuations. Industrial environments have background noise, and the second-highest historical data may itself be a noise peak. If only a difference greater than the second-place result is required (i.e., a difference > 0), it is highly susceptible to misjudging as a serious anomaly due to minor noise disturbances. Introducing a 50% confidence margin is equivalent to establishing a high signal-to-noise ratio decision window in statistics, ensuring that the highest level of response is triggered only when data quality experiences a precipitous drop, thereby guaranteeing the system's robustness under complex operating conditions.

[0050] Secondly, comprehensive risk verification is crucial. A single indicator cannot fully reflect the health status of equipment. The comprehensive status value integrates multi-source signals, which can more accurately characterize the overall risk of the equipment. This embodiment maintains a sliding window historical queue, for example, N=10. Based on the principle of low-probability events, when the current comprehensive status value ranks among the top two, it means that the probability of its occurrence has fallen outside the 20% confidence interval. This is statistically a low-probability event, indicating that the equipment status has significantly deviated from the normal state, thus judging it as an increased risk and providing an early warning. For severe risk judgment, ranking first for two consecutive periods, this design utilizes time series correlation to suppress noise. Random interference in industrial sites, such as white noise, usually does not have temporal continuity. If it maintains the highest historical ranking for two consecutive periods, the possibility of random noise is greatly ruled out, confirming the persistence and determinism of the fault. This design greatly improves the system's anti-interference capability while ensuring sensitivity, avoiding false shutdowns caused by instantaneous pulses.

[0051] Then, trend anomaly verification is performed. Numerical alarms often lag behind changes in the physical process. Trend verification aims to capture the evolution direction of physical quantities and achieve predictive maintenance. Continuous same-direction changes, with a preset anomaly threshold, such as 3 cycles, are based on the principle of inertia. Changes in physical quantities usually have inertia, and three consecutive cycles of same-direction changes indicate the initial formation of a trend. At this point, although the value has not exceeded the limit, the direction has been established, and it is judged as a trend anomaly, which can capture the nascent stage of a fault. A baseline trend reversal warning is provided, ≥5 cycles with the opposite direction. This design is based on failure physics (POF) analysis. The setting of L≥5 is to filter short-term fluctuations in the processing process, such as instantaneous jumps caused by material inhomogeneity. Only changes in the same direction with a sufficiently long duration, such as more than 5 cycles, are considered a true trend. The system calculates the number of cycles of continuous same-direction changes for each dynamic baseline value. Maintain a continuously changing periodic counter in the same direction. If the trend of the current cycle has the same sign as the trend of the previous cycle (both positive or both negative), then the counter increments. If the sign changes or the current value is zero, the counter is reset to zero. For preliminary judgment of trend anomalies, if any branch exists, the number of consecutive cycles of change in the same direction... If a preset anomaly threshold is reached, it is determined to be an abnormal trend. The preset anomaly threshold is preferably 3. Three consecutive cycles of unidirectional change indicate the initial formation of a trend. Although no serious consequences have occurred at this point, it has deviated from the steady state and requires early warning. The depth of the benchmark trend reversal is determined by defining the direction of the current dynamic benchmark value change trend of the branch: from the current cycle trend value. The sign is determined by the sign. If If the current trend is upward, then the trend is upward. If the trend is downward, then the current trend is downward. Define the historical primary trend direction and calculate the arithmetic mean based on the trend values ​​of the most recent N (N≥10) periods. .like If so, the main historical trend is upward; if If so, the main historical trend is downward. That is, the two have opposite signs and are in the same direction for a consecutive number of periods. If the direction is reversed, then the determination is made in the opposite direction. Historical trends represent the normal evolutionary pattern of the equipment; for example, as the grinding wheel wears, the benchmark grinding force typically shows a slow increase. Trend. And the current consecutive L-cycle reverse change, such as a sudden and sustained decrease in parameters. This indicates an abnormal and sudden change that deviates from normal patterns, such as a sharp drop in cutting force due to a chipped grinding wheel or workpiece slippage. Such reversals often foreshadow catastrophic failures and therefore require the highest level of early warning.

[0052] The tiered response strategy balances production efficiency and equipment safety, avoiding overreaction or underreaction. Level 1 Early Warning: This is for abnormalities or increases. It only provides a warning and does not intervene in production. This is based on the availability principle, allowing the system to self-regulate or be manually observed in the early stages, avoiding unnecessary downtime losses. Level 2 Adjustment: This is for "severe" or multiple abnormalities. The system proactively intervenes, such as adjusting the feed rate. This is based on robust control theory, performing closed-loop corrections before the risk escalates. Level 3 Emergency Stop: This is for double severity or trend reversal. This is based on functional safety standards; when an irreversible failure risk is determined, a forced shutdown must be implemented to protect equipment and personnel safety. When a Level 3 emergency stop is triggered, a snapshot of the data before and after the failure is saved. Table 1 shows the logical mapping checklist for the response actions, as follows: Table 1 The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for intelligent grinding status monitoring based on multi-sensor fusion, characterized in that, include: The system acquires multidimensional heterogeneous raw data reflecting the grinding state, geometric change features reflecting the material removal effect, and mileage coding signals in real time during the grinding operation. Based on the mileage coding signals, a unified spatial reference is constructed, and the multidimensional heterogeneous raw data and geometric change features are mapped to the spatial reference to obtain a spatiotemporally aligned dataset. The multidimensional heterogeneous raw data includes at least three-dimensional grinding force, temperature field distribution image, vibration acceleration, and visual image of the grinding area. The geometric change features include at least real-time cutting depth and real-time grinding width. Grinding drive features and multidimensional response features are extracted from each mileage grid point in the spatiotemporal aligned dataset, and real-time grinding specific energy is calculated based on the mapping relationship between the two. If any of the multidimensional response features and real-time grinding specific energy of the current mileage grid point deviates from its corresponding expected response allowable range, the grid point is marked as an anomaly. According to the anomaly marking result, the data in the anomaly grid point is removed, and the data of the remaining mileage grid points are reorganized into a valid feature sequence. The multidimensional response features include at least the equivalent grinding force value, temperature rise value, and spectral energy value, and the grinding drive features include at least the statistical features of geometric change features. An iterative processing cycle is constructed, and the adaptive reference value of each feature branch in the previous cycle is obtained. Based on the adaptive reference value, the deviation feature of the feature value of each feature branch in the current cycle relative to the adaptive reference value is determined. The grinding specific energy distribution feature in this cycle is used as a global prior constraint. Combined with the deviation feature, the weight of each feature branch in this cycle is dynamically adjusted. The result of the weighted fusion of the feature values ​​of each feature branch is represented as the comprehensive state value of this cycle. The iterative processing cycle is composed of several continuous processing units based on the data distribution characteristics of the effective feature sequence. The feature value of each feature branch represents the transient thermal shock degree, mechanical vibration instability trend and abnormal surface morphology state in the current cycle. Based on the comprehensive state value and its historical evolution trend of the current iteration processing cycle, an environmental noise observer is constructed to calculate the noise benchmark that characterizes the degree of environmental interference in real time. The noise benchmark is used to adaptively correct the benchmark values ​​of each branch feature value of the previous iteration processing cycle to obtain the adaptive benchmark values ​​of each feature branch of the current iteration processing cycle, which are used for the next iteration processing cycle. Within the current iteration cycle, multi-dimensional collaborative verification is performed based on the proportion of abnormal grid points, the comprehensive state value, and the evolution characteristics of the adaptive baseline value determined by the abnormal marking results. According to the working condition confidence level generated by the verification results, a hierarchical response strategy is triggered to realize closed-loop intelligent monitoring of grinding operations.

2. The intelligent grinding status monitoring method based on multi-sensor fusion as described in claim 1, characterized in that, The spatiotemporally aligned dataset is obtained, including: During the operation of the grinding train, a pre-scanning laser profiler mounted at the front of the rail grinding equipment passes through the current section to be ground before the grinding head, and collects and generates an initial profile data sequence of the rail before grinding. The initial profile data sequence consists of several sampling points distributed along the transverse direction of the rail. After grinding is completed, the multi-source sensor detection unit mounted on the back end of the rail grinding equipment is used to collect data on the ground section in real time, generating a rail profile data sequence after grinding, a three-dimensional grinding force signal and high-frequency vibration acceleration signal time sequence, a temperature field distribution image and a visual image time sequence of the ground area. Meanwhile, the mileage encoder integrated into the walking mechanism of the grinding equipment synchronously records the mileage pulse signal. Using the mileage pulse signal, a continuous actual mileage coordinate axis along the extension direction of the rail is constructed as a unified spatial reference. Based on the spatial reference, the data sequence of the polished rail profile is spatially aligned with the initial profile data sequence, and the corresponding values ​​at the same mileage coordinates are extracted for comparison and calculation to generate a time series of geometric change features characterizing the material removal effect. The geometric change features include at least real-time cutting depth and real-time grinding width. Based on the preset spatial resolution, a standard mileage grid with equal spacing is generated on the continuous actual mileage coordinate axis. The time series of geometric change features, three-dimensional grinding force signal, high-frequency vibration acceleration signal, temperature field distribution image and visual image time series are mapped to the corresponding standard mileage grid, and the data at each grid point are aggregated to form a spatiotemporally aligned dataset.

3. The intelligent grinding status monitoring method based on multi-sensor fusion as described in claim 2, characterized in that, Generate a time series of geometric change features characterizing the material removal effect, including: The real-time operating speed of the rail grinding equipment and the physical installation distance between the pre-scanning laser profilometer and the post-scanning laser profilometer are acquired simultaneously. The time delay compensation is calculated in real time. Time delay compensation = physical installation distance / real-time operating speed. Based on the real-time running speed and mileage accumulation algorithm, the sequence of polished rail profile data collected at time t is mapped to the current mileage coordinate on the continuous actual mileage coordinate axis to form a real-time profile data sequence. Simultaneously, by using the time delay compensation amount to backtrack and calculate, the historical acquisition time corresponding to the current mileage coordinates is obtained, and the initial profile data sequence of the rail before grinding recorded at the historical acquisition time is extracted from the cache, and associated with the current mileage coordinates to form an initial profile data sequence, thereby achieving spatiotemporal alignment under a unified mileage spatial benchmark. On the continuous actual mileage coordinate axis, the initial profile value and real-time profile value corresponding to the same mileage coordinate are obtained to form a same mileage paired dataset. Based on the same mileage paired dataset, the following feature calculation is performed: Calculate the real-time cutting depth: Based on the lateral coordinates of each sampling point in the same mileage pairing dataset, align the initial profile value and the real-time profile value laterally to obtain corresponding point pairs at the same lateral position, calculate the distance difference between each corresponding point in the vertical direction of the rail, form a distance difference sequence, remove the maximum value of the preset proportion in the sequence, select the maximum value in the remaining sequence, or use the high-order preset quantile value of the sequence as the real-time cutting depth of the rail cross section at the current mileage coordinate; Calculate the real-time grinding width: filter out sampling points whose absolute value of the distance difference exceeds the preset grinding threshold, and take them as valid sampling points. Aggregate the horizontally adjacent valid sampling points into at least one continuous grinding area. Select the main grinding area with the largest horizontal span. Based on the horizontal coordinate difference of the boundary point of the main grinding area, determine the real-time grinding width of the rail cross section at the current mileage coordinate.

4. The intelligent grinding status monitoring method based on multi-sensor fusion as described in claim 2, characterized in that, Recombined into a valid feature sequence, including: (1) For each standard mileage grid in the spatiotemporal aligned dataset, extract and calculate multidimensional response features from the multidimensional heterogeneous data associated with that grid point, wherein: The three-dimensional grinding force signal data of the grid point is statistically processed, and its mean value is defined as the equivalent grinding force value. The difference between the highest temperature value and the ambient reference temperature is extracted from the temperature field distribution image data and defined as the temperature rise value. The high-frequency vibration acceleration signal data is frequency domain transformed, and its power spectral density integral value is defined as the spectral energy value. Meanwhile, grinding drive features are extracted from the geometric change features associated with the grid point, including: calculating the average real-time cutting depth and the average real-time grinding width within the grid point. And obtain the real-time operating speed of the grinding equipment derived from the mileage encoding signal; (2) Based on the equivalent grinding force and the real-time running speed, calculate the instantaneous grinding power; based on the average real-time cutting depth, the average real-time grinding width, and the real-time running speed, calculate the instantaneous material removal rate; and based on the instantaneous grinding power and the instantaneous material removal rate, calculate the real-time grinding specific energy of the grid point. (3) Construct the measured state vector of the grid point, including the real-time grinding specific energy, temperature rise value, and spectral energy value, and perform two-level discrimination, wherein... First, determine whether the real-time grinding specific energy is within the preset reasonable range of grinding specific energy; Secondly, if it is within the reasonable range of the grinding specific energy, then according to the preset dynamic allowable range mapping relationship, find the allowable upper and lower limits corresponding to the real-time grinding specific energy, and determine whether the temperature rise value and the spectrum energy value exceed the range. (4) If any of the above judgment results are negative, the grid point and all corresponding data are marked as abnormal and the abnormal data is removed. If all the above judgment results are positive, the grid point and all corresponding data are marked as valid data. During the process of traversing all standard mileage grids, data marked as abnormal is removed in real time, and valid data is reorganized in order according to the arrangement of its corresponding standard mileage grids on the continuous actual mileage coordinate axis to form a valid feature sequence of multi-source synchronization.

5. The intelligent grinding status monitoring method based on multi-sensor fusion as described in claim 4, characterized in that, The weighted fusion of the eigenvalues ​​of each feature branch is represented as the comprehensive state value for that period, including: The iterative processing cycle is based on the effective feature sequence, which is divided into several consecutive mileage processing units according to a preset sequence step size, and the data processing process of each processing unit is mapped to a time iteration cycle of the control system; wherein, the processing unit is composed of consecutively arranged effective grid points in the effective feature sequence; Iterate through the given iteration cycles, and for the current iteration cycle, execute the following steps in real time: (1) Read the original mileage coordinates of the first and last effective grid points in the unit and calculate the physical span. Based on the physical span and the preset standard grid resolution, calculate the number of theoretical grid points that should be included in the span. Calculate the difference between the number of theoretical grid points and the actual number of effective grid points in the mileage processing unit, and determine the ratio of the difference to the number of theoretical grid points as the proportion of abnormal grid points. Based on the mileage processing unit corresponding to the current iteration processing cycle, extract the real-time grinding specific energy values ​​of all mileage grid points in the unit, and calculate the arithmetic mean of the values ​​as the representative grinding specific energy in the cycle. Based on the temperature field distribution image corresponding to each mileage grid point in the mileage processing unit, the ratio of the temperature rise difference between adjacent grid points to the mileage distance is calculated to obtain the temperature rise gradient sequence. The maximum value in the temperature rise gradient sequence is taken as the transient thermal shock branch characteristic value in this period. Extract the spectral energy values ​​of all mileage grid points within the mileage processing unit, and calculate the root mean square value of the spectral energy values ​​as the characteristic value of the mechanical vibration instability branch within that period; Color space conversion and threshold segmentation are performed on the visual image of the polishing area of ​​each mileage grid point in the mileage processing unit. The proportion of single-point thermal damage area of ​​the grid point is calculated. The proportion of single-point thermal damage area of ​​all mileage grid points in the current cycle is compared. The maximum value is selected as the thermal damage morphology anomaly degree of the current cycle. The thermal damage morphology anomaly degree is used as the surface morphology anomaly branch feature value of the current cycle. (2) Construct a multi-dimensional state observation set, execute a hierarchical weighted fusion strategy, and obtain the comprehensive state value generated by multi-source feature fusion.

6. The intelligent grinding status monitoring method based on multi-sensor fusion as described in claim 5, characterized in that, Implement a layered weighted fusion strategy, including: Based on the current iteration processing period T k Representative grinding specific energy within The linear normalization method is used to map them to global prior weight coefficients. The calculation formula is as follows: ； Where k is the index of the current iteration cycle, k = 1, 2, ..., E min and E max These are the preset lower and upper threshold values ​​for grinding specific energy under normal grinding conditions, respectively, P min and P max These are the preset lower and upper limits for the global prior weight coefficients, respectively. `clip(·)` represents the truncation function, used to restrict the calculation results to [P]. min P max Within the interval; Get the current iteration processing period T k eigenvalues ​​of each feature branch And read the previous iteration period T from memory. k-1 Updated and stored baseline values ​​for each feature branch Calculate the normalized deviation The calculation formula is as follows: ； Where i=1,2,3, i=1 represents the transient thermal shock branch, i=2 represents the mechanical vibration instability branch, and i=3 represents the morphological anomaly branch. Based on the current iteration processing period T k Global prior weight coefficients and the normalized deviation Calculate the adaptive fusion weights of the i-th branch. The calculation formula is as follows: ； Where α is the deviation response sensitivity coefficient, β is the minimum basic weight coefficient, γ is the gain coefficient, n is a nonlinear sensitivity adjustment factor greater than 1, and K th K is the preset overall energy load warning threshold. th ∈[E min E max ]; In the current iteration processing period T k Internally, the calculated weights of each branch are used for adaptive fusion. The eigenvalues ​​of the transient thermal shock branch, the mechanical vibration instability branch, and the abnormal surface morphology branch are weighted and fused to obtain a comprehensive state value. This comprehensive state value characterizes the overall load level determined by grinding specific energy, thermal shock, mechanical vibration, and surface morphology under the current iterative processing cycle.

7. The intelligent grinding status monitoring method based on multi-sensor fusion as described in claim 6, characterized in that, Obtain the adaptive baseline values ​​for each feature branch in the current iteration processing cycle, including: The current iteration processing period T is generated using a first-order exponential smoothing algorithm. k Predicted comprehensive state value The formula is as follows: ； Where α is the smoothing coefficient. The previous iteration processing period T k-1 The overall state value, Previous iteration processing cycle T k-1 The predicted comprehensive state value; Using the current iteration processing period T k Comprehensive state value Combined with predicted state values Calculate the current iteration processing period T k absolute value of prediction residuals The formula is as follows: ； Calculate the moving average of the absolute values ​​of the prediction residuals over the most recent M periods, and use it as the current iteration period T. k Dynamic noise benchmark The formula is as follows: ; Where M is the preset sliding window length, M>3, and j is the temporal index variable within the sliding window, representing the j-th cycle backtracking from the current iteration processing cycle; Calculate the current iteration processing period T k The ratio of the absolute value of the predicted residual to the dynamic noise baseline yields the relative disturbance amplitude. ,in, To prevent division by zero, the value is set to... The formula is as follows: ； Construct global scaling factor ,in, The sensitivity coefficient is given by the following formula: ； Using the global scaling factor For the previous iteration period T of each of the three feature branches k-1 benchmark value Perform synchronous updates to obtain the baseline values ​​of each feature branch updated in the current iteration cycle. The formula is as follows: ； The current iteration processing cycle benchmark value Write to memory as the next iteration period T k+1 The normalized deviation criterion is as follows: if the current period is the initial period, then the preset initial criterion value is used as the adaptive criterion value for the current period.

8. The intelligent grinding status monitoring method based on multi-sensor fusion as described in claim 7, characterized in that, Based on the operating condition confidence level generated from the verification results, a tiered response strategy is triggered, including: Within the current iteration processing cycle, key indicators are extracted, including: the proportion of abnormal grid points, the comprehensive status value, and the trend of the benchmark value. The trend of the benchmark value represents the direction of change of the benchmark value updated in the current iteration processing cycle relative to the benchmark value of the previous iteration cycle. Maintain a historical data queue for the most recent N periods, where N ≥ 10. This queue contains the aforementioned key metrics. Perform the following adaptive validation: (1) Data quality verification: The percentage of abnormal grid points in the current iteration processing cycle is sorted and compared with all values ​​in the historical queue. If the percentage of abnormal grid points in the current iteration processing cycle ranks first in the historical queue or exceeds the preset basic threshold, it is determined that the data quality is abnormal. If the abnormal grid point in the current iteration processing cycle ranks first in the historical queue and the difference between it and the second historical grid point exceeds the preset significance ratio, it is determined that the data quality is seriously abnormal. (2) Comprehensive risk verification: The comprehensive status value of the current iteration processing cycle is sorted and compared with all values ​​in the historical queue. If the comprehensive status value of the current iteration processing cycle ranks among the top two in the historical queue, it is determined that the comprehensive risk has increased. If the comprehensive status value of the current iteration processing cycle ranks first in the historical queue and remains first for two consecutive cycles, it is determined that the comprehensive risk is serious. (3) Trend anomaly verification: For each benchmark value change trend, calculate the number of consecutive cycles of change in the same direction; if there is any branch whose number of consecutive cycles of change in the same direction reaches the preset anomaly threshold, it is judged as a trend anomaly; if there is any branch whose number of consecutive cycles of change in the same direction reaches L cycles, L≥5, and the current benchmark value change trend of the branch is opposite to the historical main trend direction, it is judged as a benchmark trend reversal warning. Based on the above verification results, the following hierarchical response strategy will be implemented: Level 1 warning: When any verification item is determined to be abnormal or elevated, a warning signal is generated; Level 2 adjustment: When any verification item is determined to be serious, or when both data quality verification and comprehensive risk verification are determined to be abnormal / increased, an automatic intervention instruction is generated; Level 3 Emergency Stop: When both data quality verification and comprehensive risk verification are deemed critical, or when the baseline trend reversal warning and any critical determination occur simultaneously, an emergency stop command is generated. The generated hierarchical control commands are sent to the grinding equipment actuator to adjust the grinding operation status. At the same time, abnormal events of the current cycle are recorded in the log. If a level 3 emergency stop is triggered, the system automatically saves a number of complete data snapshots before and after the fault and reports the key information of the abnormal event to the cloud operation and maintenance platform.

Images