A Non-destructive Testing Method for Railway Slope Anchor Cables Based on Smart Sensors

Abstract

This invention discloses a non-destructive testing method for railway slope anchor cables based on intelligent sensors, comprising: acquiring multi-source time-series data and preprocessing it to form standardized input data; determining event times and event windows, and generating event marker sequences; constructing features to obtain input sample sequences; constructing an improved liquid neural network to obtain diagnostic results; performing criterion calculation and consistency verification to generate early warning trigger conditions; employing adaptive large-scale neighborhood search to obtain an updated parameter set; and performing online updates and risk classification early warnings to form evaluation results and early warning information. This invention combines an improved liquid neural network with adaptive large-scale neighborhood search parameter optimization to achieve stable and repeatable non-destructive testing and risk classification early warning under complex railway operating conditions.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology, and in particular to a nondestructive testing method for railway slope anchor cables based on intelligent sensors. Background Technology

[0002] As a key load-bearing component of the existing railway slope support system, the existing railway slope anchor cables are susceptible to the effects of periodic dynamic loads from trains, rainfall seepage and groundwater level changes, temperature and humidity fluctuations, and corrosive media erosion during long-term service. This leads to problems such as prestress attenuation, anchorage loss, loosening, wire breakage, corrosion, and localized damage propagation, ultimately affecting slope stability and traffic safety. Currently, the methods for obtaining anchor cable status at engineering sites mainly rely on manual inspection, sampling retesting, or localized instrument testing. Although some lines have deployed sensor monitoring, these methods often rely on single measurements, resulting in scattered measurement points, limited installation conditions, asynchronous sampling, long-term drift, and unstable data quality, making it difficult to establish reproducible comparative benchmarks. Strong vibrations caused by passing trains significantly alter short-time response characteristics, leading to substantial signal differences for the same anchor cable at different times and under different traffic conditions. This makes it difficult to reliably reproduce traditional judgments based on single-data points, hindering continuous, stable, and repeatable non-destructive testing and condition assessment without structural damage.

[0003] For health assessment and risk warning of multi-source time-series monitoring data, existing technologies mostly use fixed threshold judgment, empirical rules or traditional statistical models for anomaly identification, or directly input monitoring data into a single learning model for classification and regression. They lack unified identification, alignment and segmentation processing for strong disturbance events such as train passage, which can easily misjudge transient responses induced by operational vibration as structural anomalies, or mask early damage characteristics in the context of strong noise. Baseline drift caused by temperature and humidity changes and sensor aging drift often overlap with the slow decay of prestress, making it difficult to separate long-term trends, resulting in missed detections or misjudgments. Multi-source data fusion and criterion settings are often static parameters, which cannot be adaptively updated with seasonal changes, differences in operating conditions in different sections, and changes in sensor status, making it difficult to balance false alarm rates and missed alarm rates. There is a lack of quantifiable risk classification and traceable evidence chain between early warning output and maintenance treatment, making it difficult to support refined maintenance decisions.

[0004] Therefore, how to provide a non-destructive testing method for railway slope anchor cables based on intelligent sensors is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a non-destructive testing method for railway slope anchor cables based on intelligent sensors. This invention comprehensively utilizes multi-source time-series data synchronous acquisition and preprocessing, event recognition and time-frequency feature construction, improved liquid neural network diagnosis, and adaptive large-scale neighborhood search parameter optimization to complete the entire process of acquiring mechanical, vibration, and temperature and humidity data from monitoring points, forming an input sample sequence, and outputting anchor cable health index, key state estimates, and anomaly type discrimination results. After criterion calculation and consistency verification, the risk level, risk type, and early warning trigger conditions are obtained, and the entire process of online updating and hierarchical early warning output is completed. The improved liquid neural network adopts a two-stage update of continuous advancement and event transition, a dual-state unit of slow and fast change, and a second-order liquid unit of velocity state. The adaptive large-scale neighborhood search uses an operator generator and a neighborhood search tree to achieve early warning parameter optimization and update, thereby achieving stable and repeatable detection and reducing false alarms and missed alarms under the complex operating conditions of railway vibration and temperature and humidity changes.

[0006] A non-destructive testing method for railway slope anchor cables based on intelligent sensors according to an embodiment of the present invention includes:

[0007] Intelligent sensors are deployed at the monitoring points of railway slope anchor cables to collect data synchronously from multiple sources. The data is then preprocessed to form standardized input data.

[0008] Event identification is performed based on standardized input data to determine the event time and event window, and an event tag sequence is generated.

[0009] The input sample sequence is obtained by constructing time-domain features, frequency-domain features, and event window features using the event-marked sequence.

[0010] An improved liquid neural network is constructed, and the input sample sequence is segmented through continuous advancement and event transition. The state is updated based on the dual state of slow-changing state and fast-changing state. The second-order state is updated through the second-order liquid unit, and the diagnostic results including anchor cable health index, key state estimate and anomaly type discrimination are obtained.

[0011] Based on the diagnostic results, the judgment criteria are calculated and consistency is checked to obtain the anchor cable risk level and risk type, and early warning trigger conditions are generated.

[0012] An adaptive large-scale neighborhood search is used to optimize the early warning triggering conditions. The removal and reconstruction operations are performed through an operator generator structure. The target evaluation value is calculated based on the neighborhood search tree structure to obtain the updated parameter set.

[0013] Based on the anchor cable risk level, risk type, and updated parameter set, online updates and risk classification early warning outputs are performed to form anchor cable health assessment results and early warning information.

[0014] Optionally, the multi-source time-series data includes mechanical response data, vibration response data, and temperature and humidity data.

[0015] Optionally, the process of forming standardized input data includes:

[0016] Intelligent sensors are deployed at the monitoring points of the railway slope anchor cables. Sensor calibration parameters are written and the acquisition clock is synchronized. Multi-source time-series data are obtained by synchronously acquiring data according to the sampling frequency.

[0017] Time alignment and data segment processing are performed on multi-source time series data. Data from different sensor channels are resampled and aligned to a unified time axis based on timestamps. Abnormal segments are marked and removed for time periods that meet saturation or discontinuity conditions. Missing segments are generated for time periods that do not have valid sampled values.

[0018] The multi-source time-series data after data segment processing is subjected to noise filtering, temperature and humidity drift compensation and amplitude standardization. Low-pass filtering is performed on the data of each channel, and standardized input data is generated by the method that the standardized value is equal to the original value minus the mean and then divided by the standard deviation.

[0019] Optionally, generating the event tag sequence includes:

[0020] A vibration event detection sequence is generated based on standardized input data. The root mean square value within the window is calculated using a sliding window for the vibration response data. The root mean square value is equal to the square root of the sum of the squares of the sampled values ​​within the window divided by the window length, and the root mean square sequence is obtained.

[0021] The start and end of an event are determined by the root mean square sequence. When the root mean square sequence is greater than or equal to the event threshold and the duration is greater than the minimum duration, the start of the event is determined. When the root mean square sequence is less than the event threshold and the duration is greater than the minimum end duration, the end of the event is determined, and the event time is obtained.

[0022] The event window is determined based on the event start and event end. The event window consists of the pre-buffer duration before the event start, the event duration from the event start to the event end, and the post-buffer duration after the event end. The event windows are numbered based on a unified timeline to generate an event tag sequence.

[0023] Optionally, obtaining the input sample sequence includes:

[0024] The standardized input data is segmented by event tag sequence, and event data segments are extracted according to event windows and non-event data segments are extracted according to non-event time windows.

[0025] Features are constructed for event data segments and non-event data segments, including time-domain features and frequency-domain features. Time-domain features include peak value, mean, variance, kurtosis, and impact index. The impact index is equal to the maximum absolute value divided by the average absolute value. Frequency-domain features include dominant frequency, spectral peak amplitude, band energy, and band energy ratio. Band energy is equal to the sum of the squares of the amplitudes of each frequency component within the band, and the band energy ratio is equal to the target band energy divided by the total band energy.

[0026] Based on the event data segment, an event window feature is constructed and an input sample sequence is generated. The event window feature includes event duration, event intensity, and decay feature parameters. The event intensity is equal to the maximum value of the root mean square value within the event window. The time domain feature, frequency domain feature, and event window feature are combined in chronological order to form a feature vector sequence. The feature vector sequence and the event label sequence are then combined to form the input sample sequence.

[0027] Optionally, obtaining the diagnostic results, which include anchor cable health index, critical condition estimates, and anomaly type discrimination, includes:

[0028] An improved liquid neural network is constructed, which includes a two-stage update structure of continuous advancement and event transition, a dual-state unit structure of slow-changing state and fast-changing state, and a second-order liquid unit structure of velocity state.

[0029] The input sample sequence is input into the improved liquid neural network in chronological order. For the feature vector of each time step in the input sample sequence, the corresponding event marker and the time interval information of the adjacent time step are read. The feature vector is channel aligned and dimension transformed by the input mapping layer to generate slow-change state driving input and fast-change state driving input respectively. The slow-change state, fast-change state and velocity state are initialized.

[0030] Continuous progressive update is performed at non-event moments. The slow-changing state is continuously progressively updated using the input driven by the slow-changing state, and the fast-changing state and velocity state are continuously progressively updated using the input driven by the fast-changing state. The continuous progressive update process includes progressively calculating the state derivative according to the time interval information to obtain the slow-changing state, fast-changing state and velocity state at the next moment.

[0031] When the event marker indicates the event time, the event transition update is performed. The event duration, event intensity and decay feature parameters are read according to the event window features. The event window features and the feature vector corresponding to the event time are used together to perform the event transition update. Discrete transition updates are performed on the slow-changing state, fast-changing state and velocity state. The slow-changing state, fast-changing state and velocity state after the event transition are output.

[0032] The slow-changing state, fast-changing state, and velocity state obtained from continuous updates and event transition updates are decoded and calculated to output diagnostic results, which include anchor cable health index, critical state estimates, and anomaly type discrimination results.

[0033] Optionally, obtaining the anchor cable risk level and risk type, and generating early warning triggering conditions, includes:

[0034] Read the diagnostic results and calculate the rate of change of the health index and the moving average of the health index. The rate of change of the health index is equal to the difference of the health index at adjacent times divided by the time interval between adjacent times, and the moving average is equal to the sum of the health indices within the time window divided by the length of the time window.

[0035] The prestress attenuation criterion, relaxation criterion, wire breakage criterion, and corrosion criterion are calculated based on the key state estimates and the input sample sequence.

[0036] The anomaly type discrimination results are matched with each criterion to obtain a consistency mark. Based on the consistency mark, the moving average of the health index, the rate of change of the health index, and each criterion, the anchor cable risk level and risk type are determined, and early warning trigger conditions are generated. The early warning trigger conditions include the health index threshold, the rate of change threshold, and the threshold for the number of consecutive satisfactions.

[0037] Optionally, obtaining the updated parameter set includes:

[0038] Establish a set of parameters to be optimized, including sampling parameters, event recognition thresholds and early warning triggering conditions, and set the number of candidate operators generated in each iteration and the number of expansion layers of the neighborhood search tree;

[0039] The operator generator generates destruction and repair operators. The operator generator represents destruction and repair operators in the form of instruction sequences. The instruction sequences include parameter selection instructions, removal instructions, and reconstruction instructions. The parameter selection instructions determine the subset of parameters to be operated on in the parameter set. The removal instructions perform removal operations on the subset of parameters to be operated on to obtain a partial parameter set. The reconstruction instructions perform reconstruction operations on the partial parameter set to obtain a candidate parameter set.

[0040] The candidate parameter set is expanded in multiple branches based on the neighborhood search tree structure. The current parameter set is used as the root node. For each destruction operator, a corresponding partial parameter set is generated to form the first layer of branch nodes. For each first layer of branch nodes, each repair operator is combined to generate a candidate parameter set to form the second layer of branch nodes. The expansion is repeated according to the preset number of expansion layers to obtain the neighborhood search tree.

[0041] Calculate the target evaluation value for the candidate parameter set in the neighborhood search tree, and output the updated parameter set. The target evaluation value includes the false positive rate, false negative rate and computational resource consumption. Select the candidate parameter set with the best target evaluation value as the updated parameter set in order.

[0042] Optionally, the formation of anchor cable health assessment results and early warning information includes:

[0043] Read the anchor cable risk level, anchor cable risk type, and updated parameter set, and write the updated parameter set into the online processing parameter table;

[0044] Online processing is performed using the online processing parameter table, updating the sampling frequency and event window length, updating the event identification results and event tag sequence, and updating the early warning triggering judgment results according to the early warning triggering conditions;

[0045] Based on the early warning trigger determination results, risk classification early warning information is output. The risk classification early warning information includes anchor cable identification, risk level, risk type, trigger time, trigger window and corresponding diagnostic result summary, forming an anchor cable health assessment result and early warning information record.

[0046] The beneficial effects of this invention are:

[0047] This invention achieves stable and repeatable non-destructive testing under complex railway operation conditions involving vibration and temperature / humidity variations by simultaneously acquiring and standardizing anchor cable mechanical response, vibration response, and temperature / humidity data, combined with event recognition, time-domain and frequency-domain analysis, and event window feature construction. Compared to existing methods relying on single measurements or empirical thresholds, this invention employs an improved liquid neural network to continuously advance the input sample sequence and segment it into event transitions. It utilizes a dual-state approach (slow and fast changes) and a second-order update of the velocity state to output the anchor cable health index, key state estimates, and anomaly type discrimination results. This improves the consistency and anti-interference capability in identifying states such as prestress attenuation, relaxation, wire breakage, and corrosion, while reducing false alarms and missed alarms.

[0048] This invention establishes risk levels, risk types, and early warning triggering conditions through criterion calculation and consistency verification. It then employs an adaptive large-scale neighborhood search to optimize these triggering conditions. By utilizing the removal and reconstruction operations of the operator generator, along with multi-branch expansion of the neighborhood search tree and target evaluation, it achieves dynamic updates and rapid convergence of early warning parameters. This results in the output of tiered early warning information and the generation of anchor cable health assessment results and early warning records. Compared to existing statically configured early warning methods that are difficult to adapt to different line and seasonal conditions, this invention improves the adaptability and engineering deployability of the early warning strategy, and enhances the timeliness and traceability of maintenance decisions. Attached Figure Description

[0049] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0050] Figure 1 This is a flowchart of a non-destructive testing method for railway slope anchor cables based on intelligent sensors proposed in this invention;

[0051] Figure 2 This is a block diagram of an improved liquid neural network for a non-destructive testing method for railway slope anchor cables based on intelligent sensors, as proposed in this invention.

[0052] Figure 3 This is a functional diagram illustrating the adaptive large-scale neighborhood search of a non-destructive testing method for railway slope anchor cables based on intelligent sensors proposed in this invention. Detailed Implementation

[0053] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0054] refer to Figure 1 , Figure 2 and Figure 3 A non-destructive testing method for railway slope anchor cables based on intelligent sensors, comprising:

[0055] Intelligent sensors are deployed at the monitoring points of railway slope anchor cables to collect data synchronously from multiple sources. The data is then preprocessed to form standardized input data.

[0056] Event identification is performed based on standardized input data to determine the event time and event window, and an event tag sequence is generated.

[0057] The input sample sequence is obtained by constructing time-domain features, frequency-domain features, and event window features using the event-marked sequence.

[0058] An improved liquid neural network is constructed, and the input sample sequence is segmented through continuous advancement and event transition. The state is updated based on the dual state of slow-changing state and fast-changing state. The second-order state is updated through the second-order liquid unit, and the diagnostic results including anchor cable health index, key state estimate and anomaly type discrimination are obtained.

[0059] Based on the diagnostic results, the judgment criteria are calculated and consistency is checked to obtain the anchor cable risk level and risk type, and early warning trigger conditions are generated.

[0060] An adaptive large-scale neighborhood search is used to optimize the early warning triggering conditions. The removal and reconstruction operations are performed through an operator generator structure. The target evaluation value is calculated based on the neighborhood search tree structure to obtain the updated parameter set.

[0061] Based on the anchor cable risk level, risk type, and updated parameter set, online updates and risk classification early warning outputs are performed to form anchor cable health assessment results and early warning information.

[0062] In this embodiment, the multi-source time-series data includes mechanical response data, vibration response data, and temperature and humidity data.

[0063] In this embodiment, the formation of standardized input data includes:

[0064] Intelligent sensors are deployed at the monitoring points of the railway slope anchor cables. Sensor calibration parameters are written and the acquisition clock is synchronized. Multi-source time-series data are obtained by synchronously acquiring data according to the sampling frequency.

[0065] Time alignment and data segmentation are performed on multi-source time-series data. Data from different sensor channels is resampled based on timestamps and aligned to a unified time axis. Abnormal segments are marked and removed for time periods that meet saturation or discontinuity conditions. Missing segments are marked for time periods without valid sample values. Specifically, the saturation or discontinuity conditions are:

[0066] Saturation condition refers to the situation where the output of the sensor channel reaches or exceeds the upper or lower limit of the current channel range at multiple consecutive sampling points, or the output remains at the same fixed extreme value for a long time without normal fluctuation. Discontinuous condition refers to the situation where the sensor channel experiences continuous missing data, abnormally increased sampling interval, discontinuous timestamps, or non-numerical or invalid codes at multiple consecutive sampling points, causing the data stream to be interrupted.

[0067] The multi-source time-series data after data segment processing undergoes denoising filtering, temperature and humidity drift compensation, and amplitude standardization. Low-pass filtering is applied to each channel's data, and standardized input data is generated by subtracting the mean from the original value and then dividing by the standard deviation. The temperature and humidity drift compensation specifically involves:

[0068] Temperature and humidity drift compensation refers to the process of correcting the baseline shift in the output of each sensor channel caused by changes in ambient temperature and humidity using temperature and humidity data. By establishing the correspondence between channel output and temperature and humidity, the drift component caused by changes in temperature and humidity is calculated and subtracted, so that the compensated data remains comparable under different temperature and humidity conditions.

[0069] In this embodiment, generating the event marker sequence includes:

[0070] A vibration event detection sequence is generated based on standardized input data. The root mean square value within the window is calculated using a sliding window for the vibration response data. The root mean square value is equal to the square root of the sum of the squares of the sampled values ​​within the window divided by the window length, and the root mean square sequence is obtained.

[0071] The start and end of an event are determined by the root mean square sequence. When the root mean square sequence is greater than or equal to the event threshold and the duration is greater than the minimum duration, the start of the event is determined. When the root mean square sequence is less than the event threshold and the duration is greater than the minimum end duration, the end of the event is determined, thus obtaining the event time. The event threshold is specifically set to three.

[0072] The event window is determined based on the event start and end points. The event window consists of a pre-buffer period before the event start, the event duration from the event start to the event end, and a post-buffer period after the event end. The event windows are numbered using a unified timeline as a reference, generating an event tag sequence. Specifically, determining the event window involves:

[0073] The event window is determined with the event start and event end as the center. The fixed pre-buffer time before the event start is taken as the window start segment, the time from the event start to the event end is taken as the event segment, and the fixed post-buffer time after the event end is taken as the window end segment. The three segments are spliced ​​together to form a complete event window.

[0074] In this embodiment, obtaining the input sample sequence includes:

[0075] The standardized input data is segmented using an event tag sequence. Event data segments are extracted according to event windows, and non-event data segments are extracted according to non-event time windows. The segmentation process specifically involves:

[0076] Based on the timeline corresponding to the event-marked sequence, the continuous time intervals between events marked as events and the start and end times of the corresponding event windows are used as the interception boundaries. Event data segments are intercepted from the standardized input data. The continuous time intervals between event windows marked as non-events are used as non-event time windows. Non-event data segments are intercepted from the standardized input data according to the start and end times of the windows.

[0077] Features are constructed for event data segments and non-event data segments, including time-domain features and frequency-domain features. Time-domain features include peak value, mean, variance, kurtosis, and impact index. The impact index is equal to the maximum absolute value divided by the average absolute value. Frequency-domain features include dominant frequency, spectral peak amplitude, band energy, and band energy ratio. Band energy is equal to the sum of the squares of the amplitudes of each frequency component within the band, and the band energy ratio is equal to the target band energy divided by the total band energy.

[0078] An event window feature is constructed based on the event data segment, and an input sample sequence is generated. The event window feature includes event duration, event intensity, and attenuation feature parameters. The event intensity is equal to the maximum value of the root mean square value within the event window. The time-domain feature, frequency-domain feature, and event window feature are combined in chronological order to form a feature vector sequence. The feature vector sequence and the event label sequence are then combined to form the input sample sequence. The attenuation feature parameters are specifically as follows:

[0079] The decay characteristic parameters refer to the parameters of the vibration response from strong to weak after the event ends. They are extracted based on the signal or root mean square sequence of the post-buffer period after the event end, including the decay start value, decay end value, and decay time.

[0080] In this embodiment, obtaining the diagnostic results, which include the anchor cable health index, critical condition estimates, and anomaly type discrimination, includes:

[0081] An improved liquid neural network is constructed, comprising a two-stage update structure of continuous propagation and event transition, a dual-state unit structure of slow-changing and fast-changing states, and a second-order liquid unit structure of velocity states, wherein:

[0082] The improved liquid neural network is constructed by input mapping, state update, and output decoding. The state update is set up with dual state units of slow-changing state and fast-changing state to distinguish between non-event and strong disturbance events such as trains. This allows the state to evolve smoothly in non-event situations and be quickly corrected in event situations. A velocity state is configured for the fast-changing state to form a second-order liquid unit. The update process is controlled by event markers. In non-event situations, continuous updates are performed to fully characterize the dynamic process of vibration response. In event situations, the event window features and event time features are used as inputs to perform discrete updates of event transitions. Finally, the updated state is decoded and the diagnostic results are output.

[0083] The input sample sequence is fed into the improved liquid neural network in chronological order. For the feature vector at each time step in the input sample sequence, the corresponding event marker and the time interval information of adjacent time steps are read. The feature vector is then channel-aligned and dimension-transformed by the input mapping layer to generate slow-change state-driven input and fast-change state-driven input, respectively. Initial values ​​are then assigned to the slow-change state, fast-change state, and velocity state, where:

[0084] The process of generating slow-varying state-driven inputs and fast-varying state-driven inputs involves: grouping features according to their time scale and source; classifying features used to characterize long-term trends into slow-varying feature groups, which include the moving average of strain, rate of change, and low-frequency statistics within non-event time windows; and classifying features used to characterize transient responses into fast-varying feature groups, which include the root mean square maximum value within the event window, impact index, dominant frequency, spectral peak amplitude, frequency band energy and frequency band energy ratio, and attenuation characteristic parameters; and mapping the slow-varying feature groups to slow-varying state-driven inputs and the fast-varying feature groups to fast-varying state-driven inputs through linear transformation.

[0085] The initial assignment of slow-changing state, fast-changing state and velocity state is specifically as follows: at the first moment of the processing sequence, the slow-changing state and fast-changing state are set as initial state value vectors obtained by linear transformation of the feature vector of the first time window, and the velocity state is set as the initial velocity value generated by the difference value of the vibration features within the first non-event time window.

[0086] Continuous progressive updates are performed outside of event moments. The slow-changing state is continuously updated using the slowly changing state-driven input, and the fast-changing state and velocity state are updated using the fast-changing state-driven input in a second-order continuous progressive manner. The continuous progressive update process includes progressively calculating the state derivative according to time interval information to obtain the slow-changing state, fast-changing state, and velocity state at the next moment.

[0087] The method of continuously advancing and updating the slow-changing state by using the slow-changing state driving input is as follows: the slow-changing state driving input and the current slow-changing state are updated together, the state derivative of the slow-changing state is calculated, and the current slow-changing state is advanced along the direction of the state derivative according to the time interval between adjacent time moments. The slow-changing state at the next time moment is obtained by gradually advancing the state by the method that the slow-changing state at the next time moment is equal to the current slow-changing state plus the product of the state derivative and the time interval. The slow-changing state is the internal state quantity of the long-term evolution process of the anchor cable.

[0088] The method of using the fast-changing state driving input to perform second-order continuous advancement update of the fast-changing state and the velocity state is as follows: the fast-changing state driving input, the current fast-changing state, and the current velocity state are jointly input into the second-order liquid unit, the state derivative of the velocity state is calculated, and the update relationship is that the derivative of the fast-changing state with respect to time is equal to the velocity state. The velocity state is advanced according to the time interval to obtain the velocity state at the next moment, and then the fast-changing state is advanced according to the time interval to obtain the fast-changing state at the next moment. The advancement method adopts a step-by-step advancement method in which the state at the next moment is equal to the current state plus the product of the corresponding derivative and the time interval. The fast-changing state is the internal state quantity of the transient response mode of the anchor cable under the same strong disturbance of the train passing, and the velocity state is the change rate state quantity paired with the fast-changing state.

[0089] When the event marker indicates the event time, an event transition update is performed. Based on the event window features, the event duration, event intensity, and decay characteristic parameters are read. The event window features and the feature vector corresponding to the event time are then used together for the event transition update. Discrete transition updates are performed on the slow-changing state, fast-changing state, and velocity state. The slow-changing state, fast-changing state, and velocity state after the event transition are output, where:

[0090] The event transition update is specifically performed as follows:

[0091] At the event moment, the event window features and the event moment feature vector are concatenated in a fixed order to form the event input vector, and together with the slow-changing state, fast-changing state and velocity state before the event occurs, the event transition is updated. Two layers of nonlinear mapping operations are performed on the input. The first layer is that the intermediate vector is equal to the first mapping matrix multiplied by the concatenated vector and the first bias vector. The hyperbolic tangent function is performed on the intermediate vector element by element to obtain the intermediate vector.

[0092] The second layer performs three independent linear mappings on the intermediate vector: the slow increment vector is equal to the slow increment mapping matrix multiplied by the intermediate vector plus the slow increment bias vector; the fast increment vector is equal to the fast increment mapping matrix multiplied by the intermediate vector plus the fast increment bias vector; and the velocity increment vector is equal to the velocity increment mapping matrix multiplied by the intermediate vector plus the velocity increment bias vector, resulting in three transition increment vectors.

[0093] The discrete transition update is specifically performed as follows: the discrete transition update is completed once at the event time, and additive updates are performed on the slow-changing state, fast-changing state and velocity state before the event. The update relationship is that the state after the event transition is equal to the state before the event transition plus the corresponding state increment. The updated slow-changing state, fast-changing state and velocity state are used as the initial state at the next time after the event for continuous updates.

[0094] The slow-changing state, fast-changing state, and velocity state obtained from continuous updates and event transition updates are decoded and calculated to output diagnostic results. The diagnostic results include anchor cable health index, key state estimates, and anomaly type discrimination results. The decoding calculation is specifically performed as follows:

[0095] The decoding calculation performs three-way output calculations. The first path performs linear transformation and nonlinear compression on the state vector to obtain the anchor cable health index. The calculation form is that the intermediate value is equal to the health mapping matrix multiplied by the state vector plus the health bias vector. The anchor cable health index is equal to the intermediate value, which is calculated by performing a logical function on each element and taking the scalar output. The second path performs linear transformation on the state vector to obtain the critical state estimate. The critical state estimate is equal to the state estimate mapping matrix multiplied by the state vector plus the state estimate bias vector.

[0096] The third path performs a linear transformation on the state vector to obtain the category score for each anomaly type. The anomaly type discrimination result is equal to the anomaly type corresponding to the category with the largest value among the category scores. The category score is equal to the product of the category mapping matrix and the state vector plus the category bias vector.

[0097] In this embodiment, obtaining the anchor cable risk level and risk type, and generating early warning triggering conditions, includes:

[0098] Read the diagnostic results and calculate the rate of change of the health index and the moving average of the health index. The rate of change of the health index is equal to the difference of the health index at adjacent times divided by the time interval between adjacent times, and the moving average is equal to the sum of the health indices within the time window divided by the length of the time window.

[0099] Based on the key state estimates and the input sample sequence, the prestress attenuation criterion, relaxation criterion, wire breakage criterion, and corrosion criterion are calculated. Specifically, the calculation of the prestress attenuation criterion, relaxation criterion, wire breakage criterion, and corrosion criterion involves:

[0100] Within the time window, the average prestress at the beginning of the window and the average prestress at the end of the window are calculated respectively. Prestress refers to the initial tensile stress maintained along the anchor cable axis after the anchor cable is tensioned and locked. When the attenuation amplitude is not less than the preset attenuation amplitude threshold and the attenuation rate is not less than the preset attenuation rate threshold, and the preset number of consecutive satisfactions is met, the prestress attenuation criterion is determined to be valid. The attenuation amplitude is equal to the average value at the beginning of the window minus the average value at the end of the window, and the attenuation rate is equal to the attenuation amplitude divided by the duration of the time window. The preset attenuation amplitude threshold is set to 5%, and the preset attenuation rate threshold is set to 0.1% per day.

[0101] Within a time window, obtain a sequence of estimated prestress values, calculate the cumulative decrease relative to the mean at the beginning of the window within the time window. The cumulative decrease is equal to the sum of the product of the mean at the beginning of the window at each moment minus the positive part of the estimated prestress value at that moment and the time interval. Calculate the relaxation score by dividing the cumulative decrease by the product of the mean at the beginning of the window and the duration of the time window. When the relaxation score is greater than or equal to the relaxation judgment threshold, the relaxation criterion is determined to be valid. The relaxation judgment threshold is set to 0.02.

[0102] Within the event window, the impact index and the high-frequency band energy ratio are calculated. The impact index is equal to the maximum absolute value of the impact divided by the average absolute value. The high-frequency band energy ratio is equal to the target band energy divided by the full band energy. The wire breakage score is equal to the product of the impact index and the high-frequency band energy ratio. The criterion is established when the wire breakage score is greater than or equal to the wire breakage judgment threshold. The wire breakage judgment threshold is set to six.

[0103] The relative decrease ratio of the prestress estimate and the change in the frequency band energy ratio are calculated within a long time window. The relative decrease ratio is equal to the average value at the beginning of the long window minus the average value at the end of the long window and then divided by the average value at the beginning of the long window. The change in the frequency band energy ratio is equal to the absolute value of the average value at the end of the long window minus the average value at the beginning of the long window. The corrosion score is equal to the sum of the relative decrease ratio and the change in the frequency band energy ratio. The criterion is established when the corrosion score is greater than or equal to the corrosion judgment threshold. The corrosion judgment threshold is set to 0.1.

[0104] The anomaly type identification results are matched with each criterion to obtain a consistency label. Based on the consistency label, the moving average of the health index, the rate of change of the health index, and each criterion, the anchor cable risk level and risk type are determined, and early warning trigger conditions are generated. The early warning trigger conditions include a health index threshold, a rate of change threshold, and a threshold for the number of consecutive satisfactions, where:

[0105] The determination of anchor cable risk level and risk type is as follows: the anchor cable risk type is determined by the matching result of the anomaly type judgment result and the four types of criteria. When the consistency mark is passed, the type is determined according to the corresponding type of the matching criterion, which is prestress attenuation, relaxation, wire breakage or corrosion. When the consistency mark is not passed, the risk type is determined according to the anomaly type judgment result. The anchor cable risk level is determined by the combination of the comparison result of the moving average of the health index and the rate of change of the health index and whether the criteria are valid.

[0106] The health index threshold, the rate of change threshold, and the number of consecutive satisfactions threshold are set to 0.7, 0.01 per day, and three times, respectively.

[0107] In this embodiment, obtaining the updated parameter set includes:

[0108] Establish a parameter set to be optimized, which includes sampling parameters, event recognition threshold and early warning triggering conditions, and set the number of candidate operators generated in each iteration and the number of expansion layers of the neighborhood search tree. The number of candidate operators generated is set to six, and the number of expansion layers of the neighborhood search tree is set to three.

[0109] A destruction operator and a repair operator are generated through an operator generator. The operator generator represents the destruction operator and the repair operator as a sequence of instructions. The sequence of instructions includes a parameter selection instruction, a removal instruction, and a reconstruction instruction. The parameter selection instruction determines a subset of parameters to be operated on in the parameter set. The removal instruction performs a removal operation on the subset of parameters to be operated on to obtain a partial parameter set. The reconstruction instruction performs a reconstruction operation on the partial parameter set to obtain a candidate parameter set. The generation of the destruction operator and the repair operator specifically involves:

[0110] The operator generator first extracts a subset of parameters to be operated from the parameter set in a fixed order. The subset of parameters to be operated includes parameters in the sampling parameters, event identification values ​​and warning trigger conditions. The parameter selection command, removal command and reconstruction command are concatenated in sequence to form a command sequence. The removal command deletes one or more parameter items in the subset of parameters to be operated to obtain a partial parameter set. The reconstruction command completes the deleted parameter items in the partial parameter set one by one to obtain a candidate parameter set. The command sequence is used as the destruction operator and the repair operator.

[0111] The candidate parameter set is expanded through multiple branches based on the neighborhood search tree structure. The current parameter set is used as the root node. For each destruction operator, a corresponding partial parameter set is generated to form the first-level branch node. For each first-level branch node, a candidate parameter set is generated by combining it with each repair operator to form the second-level branch node. This expansion is repeated according to a preset number of expansion layers to obtain the neighborhood search tree. Specifically, the repeated expansion according to the preset number of expansion layers to obtain the neighborhood search tree involves:

[0112] The preset expansion layer is set to three layers. The current parameter set is taken as the root node. In the first round, each destruction operator is applied to the root node in turn to obtain a partial parameter set and it is taken as the first layer branch node. In the second round, each repair operator is applied to each first layer branch node to obtain a candidate parameter set and it is taken as the second layer branch node. In the third round, each second layer branch node is taken as the starting point and the destruction operator and repair operator are applied in turn to generate a new candidate parameter set and it is taken as the third layer branch node. After the three-layer expansion is completed, the neighborhood search tree is obtained.

[0113] Calculate the target evaluation value for the candidate parameter set in the neighborhood search tree, and output the updated parameter set. The target evaluation value includes the false positive rate, false negative rate, and computational resource consumption. Then, select the candidate parameter set with the optimal target evaluation value as the updated parameter set in order. The calculation of the target evaluation value specifically involves:

[0114] The false alarm rate is equal to the number of times an alert is triggered in normal samples divided by the total number of normal samples. The false negative rate is equal to the number of times an alert is not triggered in abnormal samples divided by the total number of abnormal samples. The computational resource consumption is equal to the product of the sampling frequency and the event window length, multiplied by the number of diagnostic calculations per unit time.

[0115] In this embodiment, the process of generating anchor cable health assessment results and early warning information includes:

[0116] Read the anchor cable risk level, anchor cable risk type, and updated parameter set, and write the updated parameter set into the online processing parameter table;

[0117] Online processing is performed using the online processing parameter table, updating the sampling frequency and event window length, updating the event identification results and event tag sequence, and updating the early warning triggering judgment results according to the early warning triggering conditions;

[0118] Based on the early warning trigger determination results, risk classification early warning information is output. The risk classification early warning information includes anchor cable identification, risk level, risk type, trigger time, trigger window and corresponding diagnostic result summary, forming an anchor cable health assessment result and early warning information record.

[0119] Example 1:

[0120] To verify the feasibility of this invention in practice, it was applied to a high-fill slope section of a certain railway embankment. Twenty-four anchor cables were installed along the slope as monitoring targets. The section is under high-frequency train traffic conditions, with an average of about 180 to 220 trains passing through daily. The ambient temperature varies from -2 degrees Celsius to 34 degrees Celsius, and the relative humidity varies from 35% to 98%. During the rainy season, continuous rainfall and seepage conditions occur. The on-site monitoring signals have obvious problems such as strong disturbances from train vibration, temperature and humidity drift, and intermittent missing measurements in some channels. It is necessary to achieve stable, repeatable, and non-destructive testing of prestress attenuation, relaxation, wire breakage, and corrosion risks, and to form a graded early warning system to support maintenance decisions.

[0121] Mechanical response sensors and vibration sensors are deployed at anchor cable monitoring points, along with temperature and humidity sensors. Data is collected synchronously at a unified timestamp, and time alignment, outlier removal, missing data marking, low-pass filtering, and temperature and humidity drift compensation are performed at the edge to form standardized input data. Event identification is performed based on the standardized vibration root mean square sequence to determine the start and end times of train passage events and form event windows. Data is segmented according to event markers, and time-domain, frequency-domain, and event window features are constructed to form the input sample sequence. The input sample sequence is fed into an improved liquid neural network, and continuous propagation is performed during non-event times. The update process involves performing event transition updates at each event moment, characterizing prestress attenuation and relaxation trends with slow-changing states, and transient responses under train disturbances with fast-changing states and speed states, outputting diagnostic results. Based on the diagnostic results, criterion calculations and consistency checks are performed to obtain risk levels, risk types, and early warning triggering conditions. An adaptive large-scale neighborhood search is periodically invoked to optimize and update the early warning triggering conditions. The operator generator generates removal and reconstruction operators with instruction sequences, and the neighborhood search tree is expanded in three layers to form multi-branch candidates and output the updated parameter set for online processing of parameter table updates and hierarchical early warning output.

[0122] The system operated continuously for 90 days, collecting a total of 2,160 hours of data and accumulating 17,467 event windows. For the same anchor cable, within similar speed ranges, repeated event windows showed a stable daily health index with minor fluctuations. After temperature and humidity drift compensation, the baseline shift in cross-day comparisons significantly converged. Missing segments were marked with missing markers to maintain a consistent timeline and ensure sequence continuity. After continuous rainfall during the rainy season, the system output a risk warning for two anchor cables and generated an early warning record. On-site verification revealed signs of loosening at the anchor end, and secondary tensioning was performed. Following this, the risk level decreased and stabilized over multiple consecutive processing cycles. Another anchor cable triggered a suspected wire breakage warning during multiple train event windows due to high-frequency impacts and abnormal energy ratios. Re-inspection confirmed localized steel strand damage. The time from warning triggering to re-inspection confirmation was completed within one week, meeting the requirements of the maintenance plan.

[0123] Table 1. Comparison of Indicators of Different Methods for Non-destructive Testing of Anchor Cables on Railway Slopes

[0124] index Method of the present invention Fixed threshold ARIMA LSTM GRU Transformer False alarm rate (%) 1.8 9.6 7.4 4.1 4.6 3.7 Missed Reporting Rate (%) 2.5 12.8 10.9 5.3 5.8 4.9 Event Recognition F1 0.96 0.83 0.86 0.91 0.90 0.92 HI stability (variance) 0.0009 0.0068 0.0051 0.0024 0.0027 0.0021 Prestress estimation error (kN) 5.6 18.2 15.4 8.9 9.5 8.3 Broken wire identification F1 0.93 0.71 0.74 0.86 0.85 0.87 Calculation time (ms / window) 18 6 11 27 24 35 Energy consumption (mAh / day) 210 160 175 245 238 270

[0125] As shown in Table 1, in terms of both false positive and false negative rates, the present invention outperforms the comparative methods in both key indicators. The false positive rate is 1.8%, significantly lower than the fixed threshold method's 9.6% and ARIMA's 7.4%, and also lower than LSTM, GRU, and Transformer's 4.1%, 4.6%, and 3.7%, respectively. The false negative rate is 2.5%, also significantly lower than the fixed threshold method's 12.8% and ARIMA's 10.9%, and better than LSTM, GRU, and Transformer's 5.3%, 5.8%, and 4.9%, respectively. This demonstrates that the present invention can simultaneously reduce false positives and false negatives under complex operating conditions, avoiding the trade-off between high false positives and high false negatives encountered by traditional threshold methods and some time-series models.

[0126] From the perspective of process capability indicators, this invention demonstrates superior performance in event recognition and health index stability. The event recognition F1 score reaches 0.96, higher than the fixed threshold of 0.83 and ARIMA's 0.86, and also higher than LSTM, GRU, and Transformer's 0.91, 0.90, and 0.92, respectively. The health index stability variance is 0.0009, significantly better than the fixed threshold of 0.0068 and ARIMA's 0.0051, and also better than LSTM, GRU, and Transformer's 0.0024, 0.0027, and 0.0021, respectively. This reflects that this invention provides more consistent recognition of train event windows and outputs a more stable health index, which is beneficial for long-term trend comparison and the repeatability of graded early warnings.

[0127] In terms of critical state quantity accuracy, specific defect identification, and engineering cost, this invention achieves a more balanced performance in terms of prestress estimation error, wire breakage identification, and resource consumption. The prestress estimation error is 5.6 kN, superior to the fixed threshold's 18.2 kN and ARIMA's 15.4 kN, and also superior to LSTM, GRU, and Transformer's 8.9, 9.5, and 8.3 kN, respectively. The wire breakage identification F1 score is 0.93, higher than the fixed threshold's 0.71 and ARIMA's 0.74, and also higher than LSTM, GRU, and Transformer's 0.86, 0.85, and 0.87, respectively. Regarding cost, this invention has a computation time of 18 ms / window and energy consumption of 210 mAh / day, both lower than LSTM, GRU, and Transformer, but higher than the fixed threshold and ARIMA. Overall, it demonstrates significantly improved accuracy while possessing better deployment feasibility.

[0128] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A non-destructive testing method for railway slope anchor cables based on intelligent sensors, characterized in that, include: Intelligent sensors are deployed at the monitoring points of railway slope anchor cables to collect data synchronously from multiple sources. The data is then preprocessed to form standardized input data. Event identification is performed based on standardized input data to determine the event time and event window, and an event tag sequence is generated. The input sample sequence is obtained by constructing time-domain features, frequency-domain features, and event window features using the event-marked sequence. An improved liquid neural network is constructed, and the input sample sequence is segmented through continuous advancement and event transition. The state is updated based on the dual state of slow-changing state and fast-changing state. The second-order state is updated through the second-order liquid unit, and the diagnostic results including anchor cable health index, key state estimate and anomaly type discrimination are obtained. Based on the diagnostic results, the judgment criteria are calculated and consistency is checked to obtain the anchor cable risk level and risk type, and early warning trigger conditions are generated. An adaptive large-scale neighborhood search is used to optimize the early warning triggering conditions. The removal and reconstruction operations are performed through an operator generator structure. The target evaluation value is calculated based on the neighborhood search tree structure to obtain the updated parameter set. Based on the anchor cable risk level, risk type, and updated parameter set, online updates and risk classification early warning outputs are performed to form anchor cable health assessment results and early warning information.

2. The non-destructive testing method for railway slope anchor cables based on intelligent sensors according to claim 1, characterized in that, The multi-source time-series data includes mechanical response data, vibration response data, and temperature and humidity data.

3. The non-destructive testing method for railway slope anchor cables based on intelligent sensors according to claim 1, characterized in that, The formation of standardized input data includes: Intelligent sensors are deployed at the monitoring points of the railway slope anchor cables. Sensor calibration parameters are written and the acquisition clock is synchronized. Multi-source time-series data are obtained by synchronously acquiring data according to the sampling frequency. Time alignment and data segment processing are performed on multi-source time series data. Data from different sensor channels are resampled and aligned to a unified time axis based on timestamps. Abnormal segments are marked and removed for time periods that meet saturation or discontinuity conditions. Missing segments are generated for time periods that do not have valid sampled values. The multi-source time-series data after data segment processing is subjected to noise filtering, temperature and humidity drift compensation and amplitude standardization. Low-pass filtering is performed on the data of each channel, and standardized input data is generated by the method that the standardized value is equal to the original value minus the mean and then divided by the standard deviation.

4. The non-destructive testing method for railway slope anchor cables based on intelligent sensors according to claim 1, characterized in that, The generated event tag sequence includes: A vibration event detection sequence is generated based on standardized input data. The root mean square value within the window is calculated using a sliding window for the vibration response data. The root mean square value is equal to the square root of the sum of the squares of the sampled values ​​within the window divided by the window length, and the root mean square sequence is obtained. The start and end of an event are determined by the root mean square sequence. When the root mean square sequence is greater than or equal to the event threshold and the duration is greater than the minimum duration, the start of the event is determined. When the root mean square sequence is less than the event threshold and the duration is greater than the minimum end duration, the end of the event is determined, and the event time is obtained. The event window is determined based on the event start and event end. The event window consists of the pre-buffer duration before the event start, the event duration from the event start to the event end, and the post-buffer duration after the event end. The event windows are numbered based on a unified timeline to generate an event tag sequence.

5. The non-destructive testing method for railway slope anchor cables based on intelligent sensors according to claim 1, characterized in that, The process of obtaining the input sample sequence includes: The standardized input data is segmented by event tag sequence, and event data segments are extracted according to event windows and non-event data segments are extracted according to non-event time windows. Features are constructed for event data segments and non-event data segments, including time-domain features and frequency-domain features. Time-domain features include peak value, mean, variance, kurtosis, and impact index. The impact index is equal to the maximum absolute value divided by the average absolute value. Frequency-domain features include dominant frequency, spectral peak amplitude, band energy, and band energy ratio. Band energy is equal to the sum of the squares of the amplitudes of each frequency component within the band, and the band energy ratio is equal to the target band energy divided by the total band energy. Based on the event data segment, an event window feature is constructed and an input sample sequence is generated. The event window feature includes event duration, event intensity, and decay feature parameters. The event intensity is equal to the maximum value of the root mean square value within the event window. The time domain feature, frequency domain feature, and event window feature are combined in chronological order to form a feature vector sequence. The feature vector sequence and the event label sequence are then combined to form the input sample sequence.

6. The non-destructive testing method for railway slope anchor cables based on intelligent sensors according to claim 1, characterized in that, The diagnostic results obtained include anchor cable health index, critical condition estimates, and anomaly type identification, including: An improved liquid neural network is constructed, which includes a two-stage update structure of continuous advancement and event transition, a dual-state unit structure of slow-changing state and fast-changing state, and a second-order liquid unit structure of velocity state. The input sample sequence is input into the improved liquid neural network in chronological order. For the feature vector of each time step in the input sample sequence, the corresponding event marker and the time interval information of the adjacent time step are read. The feature vector is channel aligned and dimension transformed by the input mapping layer to generate slow-change state driving input and fast-change state driving input respectively. The slow-change state, fast-change state and velocity state are initialized. Continuous progressive update is performed at non-event moments. The slow-changing state is continuously progressively updated using the input driven by the slow-changing state, and the fast-changing state and velocity state are continuously progressively updated using the input driven by the fast-changing state. The continuous progressive update process includes progressively calculating the state derivative according to the time interval information to obtain the slow-changing state, fast-changing state and velocity state at the next moment. When the event marker indicates the event time, the event transition update is performed. The event duration, event intensity and decay feature parameters are read according to the event window features. The event window features and the feature vector corresponding to the event time are used together to perform the event transition update. Discrete transition updates are performed on the slow-changing state, fast-changing state and velocity state. The slow-changing state, fast-changing state and velocity state after the event transition are output. The slow-changing state, fast-changing state, and velocity state obtained from continuous updates and event transition updates are decoded and calculated to output diagnostic results, which include anchor cable health index, critical state estimates, and anomaly type discrimination results.

7. The non-destructive testing method for railway slope anchor cables based on intelligent sensors according to claim 1, characterized in that, The process of obtaining the anchor cable risk level and risk type, and generating early warning trigger conditions, includes: Read the diagnostic results and calculate the rate of change of the health index and the moving average of the health index. The rate of change of the health index is equal to the difference of the health index at adjacent times divided by the time interval between adjacent times, and the moving average is equal to the sum of the health indices within the time window divided by the length of the time window. The prestress attenuation criterion, relaxation criterion, wire breakage criterion, and corrosion criterion are calculated based on the key state estimates and the input sample sequence. The anomaly type discrimination results are matched with each criterion to obtain a consistency mark. Based on the consistency mark, the moving average of the health index, the rate of change of the health index, and each criterion, the anchor cable risk level and risk type are determined, and early warning trigger conditions are generated. The early warning trigger conditions include the health index threshold, the rate of change threshold, and the threshold for the number of consecutive satisfactions.

8. A non-destructive testing method for railway slope anchor cables based on intelligent sensors according to claim 1, characterized in that, The updated parameter set obtained includes: Establish a set of parameters to be optimized, including sampling parameters, event recognition thresholds and early warning triggering conditions, and set the number of candidate operators generated in each iteration and the number of expansion layers of the neighborhood search tree; The operator generator generates destruction and repair operators. The operator generator represents destruction and repair operators in the form of instruction sequences. The instruction sequences include parameter selection instructions, removal instructions, and reconstruction instructions. The parameter selection instructions determine the subset of parameters to be operated on in the parameter set. The removal instructions perform removal operations on the subset of parameters to be operated on to obtain a partial parameter set. The reconstruction instructions perform reconstruction operations on the partial parameter set to obtain a candidate parameter set. The candidate parameter set is expanded in multiple branches based on the neighborhood search tree structure. The current parameter set is used as the root node. For each destruction operator, a corresponding partial parameter set is generated to form the first layer of branch nodes. For each first layer of branch nodes, each repair operator is combined to generate a candidate parameter set to form the second layer of branch nodes. The expansion is repeated according to the preset number of expansion layers to obtain the neighborhood search tree. Calculate the target evaluation value for the candidate parameter set in the neighborhood search tree, and output the updated parameter set. The target evaluation value includes the false positive rate, false negative rate and computational resource consumption. Select the candidate parameter set with the best target evaluation value as the updated parameter set in order.

9. A non-destructive testing method for railway slope anchor cables based on intelligent sensors according to claim 1, characterized in that, The formation of anchor cable health assessment results and early warning information includes: Read the anchor cable risk level, anchor cable risk type, and updated parameter set, and write the updated parameter set into the online processing parameter table; Online processing is performed using the online processing parameter table, updating the sampling frequency and event window length, updating the event identification results and event tag sequence, and updating the early warning triggering judgment results according to the early warning triggering conditions; Based on the early warning trigger determination results, risk classification early warning information is output. The risk classification early warning information includes anchor cable identification, risk level, risk type, trigger time, trigger window and corresponding diagnostic result summary, forming an anchor cable health assessment result and early warning information record.

Images