Real-time acoustic-thermal composite detection method and device for elevator steel wire rope

Abstract

The application provides an elevator steel wire rope real-time acoustic-thermal composite detection method and device, and belongs to the technical field of elevator safety detection. It solves the problem that the prior art cannot accurately reflect the overall performance state of the steel wire rope. The elevator steel wire rope real-time acoustic-thermal composite detection method comprises: multi-modal data acquisition; heterogeneous tensor construction; multi-modal data processing based on the constructed physical manifold perception and double-flow lightweight reorganization network, outputting fusion features enhanced by physical perception; based on the fusion features enhanced by physical perception, respectively outputting infrared defect prediction boxes and ultrasonic defect prediction boxes, calculating the spatial intersection-over-union ratio of the two, obtaining the fusion confidence of the steel wire rope damage based on the spatial intersection-over-union ratio; calculating the comprehensive risk index of the steel wire rope damage based on the introduced speed risk amplification factor; determining the damage risk level and outputting the grading control instruction according to the comprehensive risk index and the fusion confidence. The application can accurately determine the internal damage of the elevator steel wire rope.

Description

Technical Field

[0001] This invention belongs to the field of elevator safety inspection technology, and relates to a real-time acoustic and thermal composite detection method and device for elevator wire ropes. Background Technology

[0002] Elevator wire ropes, as an indispensable core load-bearing and power transmission component in elevator systems, run throughout the entire elevator operation process, undertaking multiple key functions. Under normal operating conditions, they not only need to stably bear the entire weight of the car, counterweight, and rated passengers or goods, ensuring balanced load transmission during vertical lifting, but also transmit power torque through the frictional coupling effect with the traction sheave. Working in conjunction with the traction machine, they achieve smooth start-up, uniform speed operation, and precise braking of the elevator. Their structural stability directly determines the smoothness and reliability of elevator operation. In emergency situations, they work closely with safety protection devices such as speed governors and safety brakes. When the car speed reaches 115% of the rated value, the speed governor will activate first, sending an electrical signal to cut off the power source; if the speed continues to increase abnormally, the safety brakes will be pulled by the wire rope, forcibly braking the car onto the guide rails to prevent a fall accident. Therefore, steel wire rope is one of the components with the highest frequency of problems in elevator inspection. Its performance directly affects the safety of elevator operation and its importance cannot be ignored. It is not only the core carrier that bears the suspension and transmits power, but also the "lifeline" that connects the speed governor and safety brake to achieve the final safe braking.

[0003] In practical applications, elevator wire ropes are prone to various failures due to factors such as working environment, stress characteristics, and maintenance level. Common problems mainly cover the following aspects: First, mechanical damage problems, including normal wear caused by long-term friction, local wear caused by uneven contact with rope pulleys or drums, deformation wear caused by external extrusion, and abnormal morphology such as kinking, wavy deformation, and local bulging. According to national standards, when the wear of the wire rope diameter exceeds 10% of the nominal diameter, it must be replaced in time. Second, corrosion problems, including surface rust induced by humid environment and lubrication failure. In addition to internal corrosion, there is also chemical corrosion caused by damage to the steel wire coating in acidic or alkaline environments. Corrosion significantly reduces the tensile strength of the steel wire rope, creating potential safety hazards. Thirdly, there are fatigue and fracture issues. The alternating stress generated by frequent elevator starts and stops can cause fatigue micro-cracks in the steel wire rope, which can then develop into localized, scattered, or internal wire breakage. When the number of broken wires within the same lay exceeds 5% of the total number of wires, the elevator must be immediately taken out of service and replaced. Furthermore, malfunctions in auxiliary components such as moldy and broken rope cores, dried-out and contaminated lubricating grease, and loose or detached joints can also exacerbate the deterioration of the steel wire rope's performance, further shortening its service life. All of these problems can lead to an imbalance in the stress characteristics of the steel wire rope, which in severe cases can cause major safety accidents such as wire rope breakage and car falls. Therefore, standardized inspection and timely maintenance are necessary to ensure the safe operation of elevators.

[0004] Currently, the core inspection methods for elevator wire ropes mainly include visual inspection, ultrasonic testing, and infrared thermal imaging. Visual inspection is simple to operate and relatively inexpensive, but it is greatly affected by subjective factors such as the professional level and sense of responsibility of the inspectors, resulting in limited accuracy. Furthermore, it cannot achieve real-time online monitoring and is difficult to detect hidden defects within the wire rope. While single-modal inspection methods such as ultrasonic testing and infrared thermal imaging can compensate for some of the shortcomings of visual inspection, they suffer from high false alarm rates and difficulties in qualitative and quantitative analysis of defects. They cannot comprehensively and accurately reflect the overall performance status of the wire rope, making it difficult to meet the high-precision requirements of elevator safety inspection. Summary of the Invention

[0005] The purpose of this invention is to address the aforementioned problems in the existing technology by proposing a real-time acoustic-thermal composite detection method and device for elevator wire ropes. The technical problem to be solved is: how to accurately determine the internal damage of elevator wire ropes.

[0006] The objective of this invention can be achieved through the following technical solution: A real-time acoustic-thermal composite detection method for elevator wire ropes, comprising the following steps:

[0007] Multimodal data acquisition steps: Acquire ultrasonic signals inside the wire rope, infrared two-dimensional temperature data on the surface of the wire rope, and pulse code signals of the wire rope operation;

[0008] The steps for constructing heterogeneous tensors include:

[0009] The ultrasonic signal is subjected to a complex Morlet wavelet continuous transform to reconstruct a two-dimensional ultrasonic time-frequency matrix, and then extended into a high-dimensional ultrasonic time-frequency tensor.

[0010] The infrared two-dimensional temperature data is subjected to adaptive statistical normalization processing and reconstructed into an infrared feature tensor;

[0011] The pulse-coded signal is solved and normalized, and a physical manifold control vector is generated through nonlinear mapping;

[0012] The constructed physical manifold sensing and dual-stream lightweight reassembly network is used to enter the multimodal data processing step, including:

[0013] Ultrasonic features and infrared features are extracted from the ultrasonic time-frequency tensor and the infrared feature tensor respectively through a dual-stream backbone network, and then spliced ​​to generate a multimodal fusion feature tensor.

[0014] In the feature fusion stage, a physical gain factor is generated based on the physical manifold control vector and fused with the multimodal fusion feature tensor to output the fused feature enhanced by physical perception.

[0015] Based on the fusion features enhanced by physical sensing, the detection head outputs infrared defect prediction boxes and ultrasonic defect prediction boxes respectively.

[0016] Decision-making and hierarchical management steps:

[0017] Execute bimodal mutual verification logic to calculate the spatial intersection-union ratio (CIU) of the infrared defect prediction box and the ultrasonic defect prediction box; obtain the fusion confidence of the wire rope damage based on the spatial CIU.

[0018] The comprehensive risk index of wire rope damage is calculated based on the introduced speed risk amplification factor;

[0019] Based on the comprehensive risk index and fusion confidence level, the damage risk level is determined, and a graded control instruction corresponding to the risk level is output.

[0020] When the wire rope traction system starts, this real-time acoustic-thermal composite detection method for elevator wire ropes begins to detect wire rope damage in real time. It acquires ultrasonic signals from inside the wire rope in real time, and reconstructs them into a two-dimensional ultrasonic time-frequency matrix through complex Morlet wavelet continuous transform. This matrix is ​​then expanded to a high-dimensional ultrasonic time-frequency tensor through channel dimension expansion to capture transient spectral characteristics. Simultaneously, it acquires infrared two-dimensional temperature data from the wire rope surface, which is then reconstructed into an infrared feature tensor after adaptive statistical normalization to capture surface thermal texture. At the same time, it acquires pulse-coded signals from the wire rope's operation in real time. After calculation and normalization, these signals are mapped to generate a high-dimensional physical manifold control vector. In other words, through nonlinear mapping, discrete physical scalars are projected onto the high-dimensional feature space of the neural network, forming a continuously changing state control signal that serves as the adjustment knob for the neural network. The three data streams are then converged into the network. A dual-stream backbone network extracts acoustic and thermal features respectively. At the feature fusion neck, the physical manifold control vector serves as the prior control signal. Through a dynamic gain gating mechanism, it automatically compensates for thermal hysteresis and nonlinear drift caused by variable speed operation at the feature level. Finally, the detection head outputs infrared and ultrasonic defect prediction boxes. In the decision-making and hierarchical control stage, a dual-modal mutual verification logic is executed: first, the spatial intersection-union ratio of the infrared and ultrasonic defect prediction boxes is calculated to eliminate false alarms caused by unilateral environmental interference; then, a comprehensive risk index is calculated based on the real-time speed risk amplification factor to quantify potential hazards under high-speed conditions; finally, control commands are output based on the risk level to determine the damage status of the elevator wire rope, achieving accurate assessment of internal damage. Furthermore, this method utilizes the energy dissipation characteristics of ultrasound, using it simultaneously as a detection signal and a thermal excitation source. During propagation, ultrasound not only reflects the acoustic integrity of the internal structure but also actively heats potential closed cracks, achieving low-power, integrated detection without the need for additional heating equipment.

[0021] In the above-mentioned real-time acoustic-thermal composite detection method for elevator wire ropes, the operation of constructing the ultrasonic time-frequency tensor includes:

[0022] First, define a complex Morlet mother wavelet basis function. :

[0023] ;

[0024] in, The time independent variable of the basis function; Indicates bandwidth parameter; Indicates the center frequency; Represents the imaginary unit; Pi is a constant. Denotes the phase quadratic term factor. This represents the acoustic wave dispersion phase compensation coefficient;

[0025] Construct physical frequency With mathematical scale The inverse mapping relationship between them:

[0026]

[0027] in, This indicates the sampling rate of the data acquisition card, which is the number of discrete data points collected per second. This represents the high-frequency attenuation adaptive gain term;

[0028] Using the above basis functions The original ultrasound signal is convolved, and the correlation strength of the signal at different time points and frequency scales is extracted using Formula 1. Formula 1 is as follows:

[0029]

[0030] in, These are complex wavelet coefficients, representing the signal at a specific time. and specific scale Similarity; Represents the original ultrasound signal; This represents the translation parameter, corresponding to the time axis coordinates of the output time-frequency plot; Represents the mother wavelet function;

[0031] The ultrasonic time-frequency tensor is calculated using Formula 2, which is:

[0032]

[0033] in, This represents the ultrasonic time-frequency tensor input to the network; This represents the modulo-square operation; Indicates the noise floor cutoff threshold; Indicates the nonlinear amplification factor; This indicates that it is used to filter out invalid background signals that are below the ambient noise level. This represents a numerical stability constant to prevent computational overflow caused by taking the logarithm of 0.

[0034] The definition of the complex Morlet mother wavelet basis function enables the precise capture of extremely weak transient defect features in ultrasonic signals (such as microsecond-level echoes caused by wire breakage). By calculating the ultrasonic time-frequency tensor using Formula 2, the one-dimensional waveform is transformed into a two-dimensional image tensor, enabling the network to simultaneously identify strong echoes (large cracks) and weak echoes (minor wear), providing a fundamental guarantee for the accurate determination of internal damage to elevator wire ropes.

[0035] In the above-mentioned real-time acoustic-thermal composite detection method for elevator wire ropes, the operation of constructing the infrared feature tensor includes:

[0036] Define a nonlinear mapping model based on the statistical moments of a local sliding window:

[0037]

[0038] in, This represents the texture-enhanced infrared feature tensor that is ultimately input into the network; This represents the raw two-dimensional infrared temperature data acquired by the infrared thermal imager; and Representing pixels Centered The mean background temperature and the standard deviation of the temperature texture within a local neighborhood window; This is a numerical stability constant to prevent the denominator from being zero; This is a contrast enhancement factor. This step effectively eliminates nonlinear lighting interference caused by localized oil stains or uneven metal reflection in the elevator shaft environment, and maximizes the highlighting of subtle defect temperature rises.

[0039] In the above-mentioned real-time acoustic-thermal composite detection method for elevator wire ropes, the operation of constructing the physical manifold control vector includes:

[0040] The pulse-coded signal includes real-time instantaneous speed and running distance. First, the real-time instantaneous speed and running distance are normalized and periodically encoded:

[0041] ;

[0042] ;

[0043] in, This represents the normalized velocity scalar; This indicates the real-time instantaneous speed of the encoder feedback; This indicates the elevator's rated maximum speed; Indicates the speed sensitivity coefficient; This represents the hyperbolic tangent activation function, which maps the velocity to the interval (-1, 1). Indicates the absolute mileage of the wire rope; Indicates the lay distance of the wire rope; Represents a positional periodic encoding vector;

[0044] An initial physical feature vector containing velocity and phase is constructed using Formula 3, and a nonlinear manifold mapping is performed. Formula 3 is as follows:

[0045] ;

[0046] ;

[0047] in, Represents the physical manifold control vector; This is the initial physical feature vector; The learnable weight matrix of the multilayer perceptron. This is the bias vector of the multilayer perceptron; This step modifies the linear unit activation function, introducing nonlinear features. This operation endows the neural network with the ability to perceive the elevator's operating status, providing a more efficient and reliable feature base for subsequent status monitoring and fault detection.

[0048] In the above-mentioned real-time acoustic-thermal composite detection method for elevator wire ropes, the operations for extracting ultrasonic and infrared features both include:

[0049] The infrared feature tensor and the ultrasonic time-frequency tensor are jointly defined as the input tensor. , Represents the infrared feature tensor. Represents the time-frequency tensor of ultrasound;

[0050] Input tensor Divide into active and passive parts:

[0051] ;

[0052] ;

[0053] in, Indicates the number of channels; This indicates the number of active channels involved in the calculation; The component involved in the calculation represents the feature component that has been segmented and is used in the actual spatial convolution calculation. This represents the remaining feature components that do not participate in the convolution operation and are directly subjected to identity mapping;

[0054] Formula 4 is used to pair the eigencomponents of the active part. For feature extraction, Formula 4 is as follows:

[0055] ;

[0056] ;

[0057] in, These are the extracted active space features; Output feature map; for Convolution kernel weight matrix; For texture direction prior operators; This represents the bias term of the convolutional layer; It is a learnable channel scaling factor used to adaptively adjust the feature amplitude to enhance weak defect signals.

[0058] In the above-mentioned real-time acoustic-thermal composite detection method for elevator wire ropes, the operation of generating the multimodal fusion feature tensor includes:

[0059] Introducing perceptual weights The infrared / ultrasound multi-scale aggregated feature tensor is output using a pyramid pooling module for weighted fusion. The calculation formula is as follows:

[0060] ;

[0061] ;

[0062] In the formula, For infrared / ultrasound multiscale aggregated feature tensor; For the first Hierarchical pooling characteristics; It is the max pooling operator; Represents the convolution kernel; The values ​​are 1, 2, and 3. This indicates a weighted concatenation based on the channel dimension; The scale importance coefficients are automatically learned by the network during backpropagation;

[0063] The output infrared multiscale aggregated feature tensor and ultrasonic multiscale aggregate feature tensor The data is concatenated along the channel dimension to generate a multimodal fusion feature tensor. .

[0064] In the aforementioned real-time acoustic-thermal composite detection method for elevator wire ropes, the operations for outputting the fused features enhanced by physical sensing include:

[0065] Employing a large convolutional kernel decomposition strategy to fuse feature tensors from multiple modalities Extract long-range spatial dependency features, namely:

[0066] ;

[0067] ;

[0068] in, This is an intermediate value representing the horizontal characteristic. This represents the final long-range spatial dependency feature; This represents depthwise separable convolution; Indicates the size of the large convolution kernel; Represents asymmetric convolution kernels in the horizontal and vertical directions;

[0069] The physical gain factor is generated based on the physical manifold control vector, and the operation is as follows:

[0070]

[0071] ;

[0072] in, Indicates a velocity-dependent offset; Represents the physical manifold control vector; Indicates the weights and biases of the gated projection layer; Physical gain factor;

[0073] Based on the above physical gain factor We construct horizontal attention and velocity-modulated vertical attention separately, thereby outputting fused features enhanced by physical perception:

[0074]

[0075] ;

[0076] ;

[0077] in, This indicates a fusion feature enhanced by physical perception. Indicates the use of channel information fusion convolution; The temperature coefficient for attention sharpening is introduced; Sigmoid represents the activation function. Indicates horizontal attention weights; This represents the vertical attention weight. Through the design of multimodal data aggregation and multi-scale physical injection, it is ensured that the corresponding feature tensors, whether for tiny broken wires or long-distance wear zones, can receive velocity gain modulation that matches their receptive field during the multi-scale fusion stage. This completely eliminates the spatiotemporal misalignment error caused by thermal hysteresis under variable frequency conditions across the entire scale range, improving detection accuracy.

[0078] In the above-mentioned real-time acoustic-thermal composite detection method for elevator wire ropes, the fusion confidence level of wire rope damage is obtained using Formula 5, which is:

[0079] ;

[0080] In the formula, Indicates the first The fusion confidence of the objectives; This represents the raw confidence level of the infrared and ultrasonic branches output by the network; The spatial intersection-union ratio (IUU) of the infrared defect prediction box and the ultrasonic defect prediction box is based on the real-time decoding output. The absolute coordinate vector of the defect prediction bounding box for each target and Dynamically calculated geometric verification variables; spatial intersection-union ratio obtained through Formula 6. :

[0081] ;

[0082] In the formula, Indicates the area to be calculated. and Representing the corresponding number The method calculates the spatial overlap and total coverage area of ​​the acoustic and thermal modal prediction boxes for a target. When the acoustic and thermal modal prediction boxes are highly overlapping in space (i.e., high IoU), the fusion confidence score approaches 1. Conversely, if the spatial IoU is low, the detected signal is determined to be a non-homogeneous spurious artifact. This method can automatically eliminate such false alarms by calculating the fusion confidence score, thereby ensuring the physical authenticity of the detection results.

[0083] In the aforementioned real-time acoustic and thermal composite detection method for elevator wire ropes, the comprehensive risk index is calculated using Formula 7. :

[0084] ;

[0085] In the formula, The pixel area of ​​the defect box; Indicates the first The fusion confidence of the objectives; This represents the total effective area of ​​the wire rope within the current field of view; As a speed risk amplification factor, among which This represents the risk coefficient.

[0086] In the aforementioned real-time acoustic-thermal composite detection method for elevator wire ropes, the construction of the physical manifold sensing and dual-flow lightweight reconfiguration network further includes obtaining the total loss function through Formula 8, which is:

[0087] ;

[0088] in, Represents the total loss function; The weighting factor representing the balance of losses in each component; Indicates the crossover and union ratio loss; Represents the SIoU loss function; This represents the improved classification loss value; This represents the improved regression distribution loss; This represents the final physical consistency loss value. A total loss function is designed, incorporating a velocity-weighted physical consistency loss value. During the training phase, the network autonomously learns the spatial alignment of acoustic and thermal features under high-speed conditions. This dual-modal mutual verification mechanism significantly reduces the false alarm rate caused by unilateral interference, effectively improving the accuracy of reconstructed network construction and providing a reliable foundation for the accurate detection of wire rope damage.

[0089] A real-time acoustic and thermal composite detection device for elevator wire ropes includes:

[0090] An ultrasonic transducer is used to inject ultrasonic waves into a steel wire rope and receive the reflected ultrasonic signals.

[0091] Infrared thermal imager is used to collect two-dimensional infrared temperature data of the surface of steel wire rope;

[0092] A rotary encoder is used to acquire pulse-coded signals of the wire rope in real time.

[0093] The data processing terminal is used to perform real-time tensor fusion and decision-making on ultrasonic signals, infrared two-dimensional temperature data, and pulse code signals based on its integrated physical manifold sensing and dual-stream lightweight recombination network algorithm, so as to realize the identification and graded alarm of wire rope damage.

[0094] The ultrasonic transducer, infrared thermal imager, and rotary encoder are all connected to the data processing terminal.

[0095] When the wire rope traction system is started, the elevator's real-time acoustic-thermal composite detection device begins to detect wire rope damage in real time. The ultrasonic transducer collects the voltage sequence signal inside the wire rope in real time, and reconstructs it into a two-dimensional ultrasonic time-frequency matrix through complex Morlet wavelet continuous transform. Then, it is constructed into a high-dimensional feature tensor through channel dimension expansion to capture transient spectrum features. The infrared thermal imager simultaneously collects infrared two-dimensional temperature data, and then reconstructs the original two-dimensional temperature matrix into an infrared feature tensor after adaptive statistical normalization to capture surface thermal texture. At the same time, the rotary encoder feeds back pulse code signals in real time. After solving and normalizing, it is mapped to generate a high-dimensional physical manifold control vector. The discrete physical scalar is projected into the high-dimensional feature space of the neural network, forming a continuously changing state control signal as the adjustment knob of the neural network. Subsequently, the data processing terminal, based on its integrated physical manifold sensing and dual-stream lightweight recombination network algorithm, processes the three data streams, namely, performing real-time tensor fusion and decision-making on ultrasonic signals, infrared two-dimensional temperature data, and pulse code signals, to achieve the identification and graded alarm of wire rope damage. Through the application of physical manifold sensing and dual-stream lightweight recombination network algorithm, this device achieves accurate judgment of internal damage such as broken wires, wear, and fatigue cracks in elevator wire ropes.

[0096] Compared with existing technologies, this real-time acoustic-thermal composite detection method and device for elevator wire ropes has the following advantages:

[0097] 1. This invention is based on the dual effect of ultrasonic "detection-heating". It establishes dynamic dry coupling contact between the elevator wire rope and a polyurethane acoustic roller. The continuous injection of high-power ultrasonic waves serves two purposes: firstly, as an acoustic detection wave to monitor the echo and attenuation; and secondly, as an active thermal excitation source to induce frictional heat at the defect. This eliminates the need for the induction heating coil required by traditional infrared detection, which not only simplifies the hardware structure but also significantly improves the detection sensitivity of internal defects in the wire rope. Combined with infrared thermal imager to acquire thermal images, and using physical manifold perception and dual-flow lightweight recombination network algorithm, it achieves accurate determination of internal damage such as broken wires, wear, and fatigue cracks in the elevator wire rope.

[0098] 2. This invention utilizes the attention weights of a dynamically modulated network based on a physical manifold control vector to achieve adaptive compensation for unsteady thermal hysteresis during the feature extraction stage. Building upon this, it employs bimodal spatial verification, using the intersection-union ratio of the infrared and ultrasonic defect prediction boxes as a strong verification condition to automatically eliminate unilateral artifacts. The fusion confidence level, achieved through strong acoustic-thermal coupling, ensures zero missed detections of major safety hazards while significantly suppressing false alarms caused by unilateral environmental interference. Simultaneously, it introduces an exponentially increasing velocity risk amplification factor, mapping static defect sizes to dynamic hazard indices, enabling the system to possess higher safety sensitivity under harsh operating conditions and further improving the accuracy of wire rope damage assessment. Attached Figure Description

[0099] Figure 1 This is the control flowchart of the present invention.

[0100] Figure 2 This is a schematic diagram of the physical manifold sensing and dual-flow lightweight reconfiguration network of the present invention.

[0101] Figure 3 This is a schematic diagram of the control structure of the present invention.

[0102] In the diagram, 1 is an ultrasonic transducer; 2 is an infrared thermal imager; 3 is a rotary encoder; and 4 is a data processing terminal. Detailed Implementation

[0103] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0104] This real-time acoustic-thermal composite detection method for elevator wire ropes is based on a real-time acoustic-thermal composite detection device for elevator wire ropes, such as... Figure 3 As shown, the real-time acoustic-thermal composite detection device for elevator wire ropes includes an ultrasonic transducer 1 installed inside a polyurethane acoustic roller for injecting ultrasonic waves into the wire rope and receiving the reflected ultrasonic signals; a rotary encoder 3 coaxially mounted on the side of the roller for real-time acquisition of pulse-coded signals from the wire rope's movement; an infrared thermal imager 2 installed downstream of the roller for acquiring infrared two-dimensional temperature data of the wire rope surface; and a data processing terminal 4 for real-time tensor fusion and decision-making of ultrasonic signals, infrared two-dimensional temperature data, and pulse-coded signals based on physical manifold perception and a dual-stream lightweight recombination network algorithm, thereby achieving accurate identification and graded alarm for wire rope damage such as broken wires and wear. The data processing terminal 4 includes, but is not limited to, one or more of the following: industrial computers, embedded processors, host computers, servers, mobile terminals, and dedicated detection equipment.

[0105] At the start of the test, the polyurethane acoustic rollers are pressed tightly against the surface of the running steel wire rope, establishing a dynamic dry coupling contact. For example... Figure 1 , 2 As shown, the multimodal data acquisition step is initiated: simultaneously acquiring ultrasonic signals inside the wire rope, infrared two-dimensional temperature data of the wire rope surface, and pulse code signals of the wire rope operation;

[0106] The next step involves constructing a heterogeneous tensor, including: performing a complex Morlet wavelet continuous transform on the ultrasonic signal to reconstruct a two-dimensional ultrasonic time-frequency matrix, and expanding it into a high-dimensional ultrasonic feature tensor; and performing adaptive statistical normalization on the two-dimensional infrared temperature data to reconstruct an infrared feature tensor. This achieves "heterogeneous isomorphism" in a unified high-dimensional space, laying the data foundation for subsequent deep feature fusion. The pulse-coded signal is then solved and normalized, and a physical manifold control vector is generated through nonlinear mapping.

[0107] The operations for constructing the ultrasonic feature tensor of the ultrasonic signal include:

[0108] To accurately capture extremely weak transient defect features in ultrasonic pulse signals (such as microsecond-level echoes caused by wire breakage), a complex Morlet mother wavelet basis function with optimal localization properties in both the time and frequency domains is first defined. :

[0109] ;

[0110] in, The time independent variable of the basis function; This represents the bandwidth parameter, which mainly controls the decay rate of the wavelet in the time domain. The smaller the value, the higher the time resolution, which is more conducive to accurately locating the axial position of defects during high-speed operation; This represents the center frequency, which corresponds to the main response frequency of ultrasonic transducer 1, in order to maximize the resonance response to the defect echo. To introduce the imaginary unit. The aim is to preserve the phase information of the signal and avoid feature loss caused by using only the real part; Pi is a constant. The phase quadratic term factor, The phase compensation coefficient for acoustic wave dispersion is used because the steel wire rope is a complex waveguide structure made of multiple strands of steel wire twisted in a spiral. When ultrasonic waves propagate in it, the phase velocities of different frequency components are inconsistent (i.e., dispersion effect). This compensation term aims to mathematically counteract the wave packet broadening caused by dispersion, thereby significantly improving the positioning accuracy of defect echoes in the time domain.

[0111] Next, to establish the mapping relationship between physical frequencies and mathematical scales, ensuring that the frequency axis of the time-frequency plot has a clear physical meaning, the physical frequencies are constructed. With mathematical scale The inverse mapping relationship between them:

[0112]

[0113] in, This represents the sampling rate of the data acquisition card, i.e., the number of discrete data points acquired per second. Furthermore, considering the extremely rapid attenuation of high-frequency ultrasound in the steel wire rope medium, this formula introduces a high-frequency attenuation adaptive gain term. This term is a frequency-dependent variable. A monotonically increasing nonlinear function is used to automatically compensate for energy loss in the high-frequency band when constructing the scale map, preventing the high-frequency characteristic signals of tiny broken wires from being overwhelmed by noise.

[0114] Based on this, using the aforementioned basis functions The original ultrasound signal is convolved, and the correlation strength of the signal at different time points and frequency scales is extracted using Formula 1. Formula 1 is as follows:

[0115]

[0116] in, These are complex wavelet coefficients, representing the signal at a specific time. and specific scale Similarity; Represents the original ultrasound signal; This represents the translation parameter, corresponding to the time axis coordinates of the output time-frequency plot; This represents the mother wavelet function, and the energy normalization factor ensures the energy conservation of the wavelet basis at different scales. This represents the index sequence of discrete-time sampling points.

[0117] To transform a one-dimensional waveform into a two-dimensional image tensor and compress the dynamic range through logarithmic transformation, enabling the neural network to simultaneously identify strong echoes (large cracks) and weak echoes (minor wear), the following formula two is used:

[0118] ;

[0119] in, This represents the ultrasonic time-frequency tensor input to the network; This indicates the modulo-square operation, mainly used to extract the energy (power) value of complex coefficients; This represents the numerical stability constant to prevent computational overflow caused by taking the logarithm of 0. To further improve the signal-to-noise ratio of the feature tensor, this formula introduces a noise floor truncation threshold before the logarithmic transformation. With nonlinear amplification factor , This is used to nonlinearly stretch and enhance effective defect peaks that exceed a threshold, resulting in a time-frequency tensor with higher contrast, which is beneficial for feature extraction by subsequent neural networks. This is used to filter out invalid background signals that are below the ambient noise level.

[0120] The operations for constructing infrared feature tensors include:

[0121] For the infrared measurement channel, although the data acquired by the infrared thermal imager 2 naturally possesses two-dimensional spatial attributes, the original radiation thermal map is affected by the airflow disturbance in the wellbore and the uneven emissivity of the oil on the surface of the wire rope, resulting in significant background noise and dynamic range drift. Directly inputting this data into the network would lead to instability in feature extraction. Therefore, this method first constructs an adaptive infrared input tensor based on statistical moment correction to achieve standardization and enhancement of thermal features.

[0122] To effectively eliminate nonlinear illumination interference caused by localized oil contamination or uneven metal reflection in the wellbore environment, and to maximize the highlighting of subtle defect temperature rises, this method abandons the traditional global normalization method and redefines a nonlinear mapping model based on local sliding window statistical moments:

[0123]

[0124] in, This represents the texture-enhanced infrared feature tensor that is ultimately input into the neural network. Its value is constrained to the (0, 1) interval by the Sigmoid function, which preserves the relative temperature difference features and avoids the overflow of the highlight values. This represents the raw two-dimensional infrared temperature data collected by infrared thermal imager 2, which characterizes the absolute radiation temperature distribution on the surface of the wire rope. and Representing pixels Centered The mean background temperature and standard deviation of temperature texture within the local neighborhood window can be adaptively filtered out by calculating local statistics rather than global statistics, thus removing low-frequency background fluctuations caused by uneven oil thickness on the surface of the wire rope and retaining only abnormal temperature rise signals relative to the local background. The numerical stability constant (values) ), to prevent the denominator from being zero; This is a contrast enhancement factor used to further amplify the salience of the thermal signal in the defect area.

[0125] The operations for constructing physical manifold control vectors include:

[0126] To endow the neural network with the ability to perceive the elevator's operating status, the physical scalar feedback from the rotary encoder 3 needs to be mapped into high-dimensional features. First, the pulse-coded signal is solved and normalized, that is, the real-time instantaneous speed and absolute mileage are normalized and periodically encoded:

[0127] ;

[0128] ;

[0129] in, This represents the normalized velocity scalar; This represents the real-time instantaneous velocity (m / s) of the encoder feedback. This represents the elevator's rated maximum speed and serves as the benchmark for normalization. This represents the speed sensitivity coefficient, used for adjustment. The linear response interval of the function The hyperbolic tangent activation function is used to map the velocity to the (-1, 1) interval. Indicates the absolute running mileage (mm) of the wire rope; Indicates the lay distance of the wire rope; This represents a positional periodic encoding vector.

[0130] To achieve the concatenation of physical state vectors, an initial physical feature vector containing velocity (affecting thermal hysteresis) and phase (affecting structural noise) is constructed, and a nonlinear manifold mapping is performed. This is achieved using Equation 3:

[0131] ;

[0132] ;

[0133] in, For the initial physical feature vector, This represents the physical manifold control vector, which is the "control signal" injected into the neural network and contains high-dimensional features of velocity and position; The learnable weight matrix of the multilayer perceptron. This is the bias vector of the multilayer perceptron; This indicates a modified linear unit activation function that introduces nonlinear features.

[0134] To address the limitation of computing power in elevator shaft edge computing devices, a pre-constructed physical manifold perception and dual-stream lightweight reconfiguration network is constructed. This network comprises two parallel but weighted dual-stream backbone networks, used for parallel extraction of infrared and ultrasonic features, respectively. To clarify the data flow direction, let... This represents the input feature tensor of each module in the two-stream backbone network, where the modality superscript... This is used to strictly distinguish between the infrared and ultrasonic branches. At the network's front end, the initial inputs to the infrared backbone and the ultrasonic backbone are the infrared feature tensors constructed above. With ultrasound time-frequency tensor This method utilizes the redundancy between feature map channels to transform the input feature tensor It is divided into an active part and a passive part, namely:

[0135] ;

[0136]

[0137] in, This represents the input feature tensor of each module in the two-stream backbone network. Indicates the number of channels; This represents the number of active channels participating in the calculation, typically taken as... ; The component that participates in the calculation represents the feature component that has been segmented out to participate in the actual spatial convolution calculation. This represents the remaining feature components that do not participate in the convolution operation and are directly subjected to identity mapping.

[0138] Formula 4 is used only for the eigencomponents of the active part. Feature extraction is performed to capture local texture information, and while restoring the original channel dimensions, the real-time requirements of the elevator shaft edge equipment are met. Formula 4 is:

[0139] ;

[0140] ;

[0141] in, for Convolution kernel weight matrix; To adapt to the unique helical twisted texture structure of steel wire rope, this method introduces a texture direction prior operator. The operator consists of a set of direction angles. (i.e., wire rope twist angle) Gabor filter is constructed, and the Hadamard product is performed with a standard convolution kernel. The operation gives the convolution operation an anisotropic texture perception capability, thereby effectively suppressing non-axial background noise interference; This represents the bias term of the convolutional layer; This is a learnable channel scaling factor used to adaptively adjust the feature amplitude to enhance weak defect signals; This represents the extracted active spatial features; The output feature map of the C2f-PConv module reduces computation by approximately 75% compared to standard convolution.

[0142] To construct scale-adaptive receptive fields at the ends of the backbone network, this invention improves upon the traditional fixed-kernel pooling structure by proposing a visual flow context pyramid module. Unlike existing techniques that use serially connected convolutional kernels of fixed sizes, this method employs dynamically sized convolutional kernels. Parallel or cascaded feature abstraction is performed to capture texture details at different frequencies. Furthermore, to address the feature channel redundancy and semantic dilution issues caused by simple stitching, this method introduces learnable perceptual weights. The model reconstructs output features using a weighted fusion strategy. By assigning differentiated weights to features at different scales, the model can suppress background noise (low weights) and highlight structural features of key targets such as broken fibers (high weights). The operations include:

[0143] ;

[0144] ;

[0145] In the formula, For the first Hierarchical pooling characteristics; It is the max pooling operator; Represents the convolution kernel; The values ​​are 1, 2, and 3, respectively corresponding to (Microscopic) (Middle level) (Macro) sensory field window; This indicates a weighted concatenation based on the channel dimension; The scale importance coefficients are automatically learned by the network during backpropagation; This refers to the infrared or ultrasonic multiscale aggregated feature tensor obtained after processing by the pyramid module.

[0146] Before reaching the feature fusion neckline, deep features from the infrared and ultrasonic backbones need to be converged from multiple sources. Specifically, the infrared feature multi-scale aggregation tensor output by the visual flow context pyramid module of the infrared and ultrasonic branches... With ultrasound multiscale aggregated feature tensor Rigid splicing will be performed along the channel dimension to generate a multimodal fusion feature tensor. The multimodal fusion feature tensor It will be directly used as the basic input for the Phy-LSKA module of the neck network for subsequent physical gain modulation and feature reconstruction.

[0147] To fully cover the elongated defects (cracks, broken wires) and their thermal tailing regions in wire ropes, a large convolutional kernel decomposition strategy is employed to extract long-range spatial dependency features, namely:

[0148] ;

[0149] ;

[0150] in, This indicates a depthwise separable convolution, which operates independently on each channel. The kernel size is large, corresponding to a pixel width that can cover a complete damaged area (crack + thermal tail) of the steel wire rope; Represents asymmetric convolution kernels in the horizontal and vertical directions; This is an intermediate value representing the horizontal characteristic. This represents the final long-range spatial dependency feature.

[0151] To achieve adaptive compensation, a mapping relationship from "physical velocity" to "feature gain" is established, and a physical gain factor is generated. The operation is as follows:

[0152]

[0153] ;

[0154] in, This is the control vector for the physical manifold; Indicates the weights and biases of the gated projection layer; Indicates a velocity-dependent offset; This is the physical gain factor; as the speed increases, this factor... This is used to amplify the attention weight in the vertical direction and compensate for signal blurring caused by thermal hysteresis.

[0155] Based on the above physical gain factor We construct horizontal attention and velocity-modulated vertical attention separately, and finally output a feature tensor enhanced by physical perception. The implementation method is as follows:

[0156]

[0157] ;

[0158] ;

[0159] in, Indicates the use of channel information fusion convolution. The attention sharpening temperature coefficient (value) is introduced. Here, the Sigmoid activation function does more than simply map numerical values ​​to... The probability weights are more important in the temperature coefficient. Modulation achieves a shift from a 'soft threshold' to a 'hard threshold' characteristic. That is, a smaller temperature coefficient makes the activation curve steeper, forcing attention weights to... and It tends towards a binarized distribution. This mechanism can more decisively suppress background noise and highlight locked defect features (weights approaching 1). In particular, Further affected by physical gain Multiplication modulation ( (For Hadamard product), it achieves directional amplification of the thermal tail directional features. The feature tensor, which has been enhanced by both physical sensing and temperature sharpening, i.e., the fused feature enhanced by physical sensing, is output to the detection head.

[0160] In the multi-source data stitching stage, to achieve cross-modal and multi-scale information fusion, the input of each stitching module includes a rigid aggregation of multiple heterogeneous data streams. It not only receives upsampled or downsampled feature maps from a single level of the neck network, but also synchronously receives output features from C2f-PConv modules at corresponding spatial scales in the infrared and ultrasonic backbones via lateral connections. Through this multi-source stitching, the network simultaneously aggregates deep acoustic semantics, thermal surface textures, and rich high-resolution spatial positioning information at a single processing node.

[0161] In the physical feature fusion stage, three independent Phy-LSKA modules were deployed at the high, middle, and low layers of the network, respectively. The feature tensor fused by the splicing module... The vectors are fed as inputs into the corresponding Phy-LSKA modules at their respective levels; simultaneously, these three Phy-LSKA modules receive the same high-dimensional physical manifold control vector from the front-end physical branch in parallel. Because the elevator's real-time operating speed produces a consistent physical hysteresis effect on defects of different sizes, therefore Here, it acts as a cross-scale 'global prior gating signal'. It not only informs the network of its current physical condition in advance (a priori), but also directly modulates the attention weights of each level in the form of multiplicative gating by converting them into physical gain factors.

[0162] Through the design of multi-point data aggregation and multi-scale physical injection, it is ensured that the corresponding feature tensors, whether it is a small broken wire or a long wear zone, can receive velocity gain modulation that matches its receptive field during the multi-scale fusion stage, thereby completely eliminating the spatiotemporal misalignment error caused by thermal hysteresis under frequency conversion conditions in the entire scale range.

[0163] To address the extremely small fracture surface area of ​​the broken wire in the steel wire rope and to obtain steeper gradients in the early stages of training the physical manifold perception and dual-flow lightweight recombination network, the geometric parameters of the defect prediction bounding box are first analyzed, and smaller "auxiliary inner boxes" are generated inside both the ground truth bounding box and the defect prediction bounding box. The operation is as follows:

[0164] ; ;

[0165] , ;

[0166] , ;

[0167] in, These are the center coordinates and width and height of the defect prediction bounding box output by the detection head, respectively. Represents the analytically derived geometric parameters; The shrinkage ratio factor is 0.7-0.8. This indicates the coordinates of the left / right / top / bottom boundaries of the auxiliary inner frame.

[0168] The cross-union ratio (CUI) loss is calculated based on the auxiliary inner frame, and the operation is as follows:

[0169]

[0170] ;

[0171] in, Indicates the crossover and union ratio loss; Auxiliary inner box representing the defect prediction bounding box; An auxiliary inner frame representing the actual annotation box; Center-scale operation; This indicates the area being calculated.

[0172] The predicted bounding boxes exhibit offset during regression. To significantly reduce this "oscillation" and facilitate faster and more accurate training convergence, a strategy of "aligning the axis first, then regressing the distance" is adopted, introducing an SIoU angle penalty. First, the offset components and Euclidean distance between the predicted center and the ground truth center are calculated:

[0173] , ;

[0174] ;

[0175] In the formula, This represents the absolute distance between the predicted center and the true center in the horizontal and vertical directions. The coordinates of the center of the predicted bounding box and the ground truth bounding box; This represents the straight-line Euclidean distance between two center points; To prevent division by zero of small constants.

[0176] Considering the vertically elongated geometry of elevator wire ropes within the shaft, the traditional SIoU loss function imposes equal penalties on regression errors in both the horizontal and vertical directions, which does not meet practical inspection requirements. Therefore, in wire rope inspection, the horizontal offset of the prediction box (detachment from the wire rope itself) causes a greater error than the vertical offset.

[0177] Therefore, this method improves the angle penalty term by incorporating a vertical bias. A vertically sensitive angle penalty function is constructed by introducing a vertical regression bias weight. This breaks the symmetry of the penalty, ultimately yielding the SIoU loss function. The specific operation is as follows:

[0178] ;

[0179] ;

[0180] in, This is the offset angle; To maximize the angle penalty in the horizontal direction for normalization; The angle cost coefficient; Cost of shape; As a result of distance, its decay rate is affected Dynamic modulation; These represent the network-predicted bounding boxes and the ground truth bounding boxes, respectively.

[0181] To address the issue that acoustic-thermal alignment becomes increasingly difficult and prone to errors at higher elevator speeds, a speed-weighted physical consistency loss is designed. First, the error is calculated to quantify the spatial inconsistency between the two modal detection results:

[0182]

[0183] in, Indicates the center coordinates of the infrared branch prediction box; This indicates the center coordinates of the ultrasound branch prediction box.

[0184] To achieve an exponential increase in weights as the elevator speed increases, the network is forced to focus more on alignment under high-speed (severe thermal hysteresis) conditions, resulting in speed-dependent penalty weight calculations.

[0185] ;

[0186] in, Indicates the dynamic physical penalty weight; Indicates the maximum rated speed; This is the stringency factor (hyperparameter).

[0187] The final physical consistency loss value is:

[0188] ;

[0189] To adapt to the special working conditions of elevator wire rope inspection, this method does not directly adopt the loss function of general object detection, but makes adaptive improvements for high-speed ambiguity and slender geometric features. To address the risk of missed detections caused by blurred defect features under high-speed operation, and the class imbalance problem due to the extreme scarcity of fault samples, this method proposes a speed-aware focus loss. This loss function introduces a speed gain coefficient on top of Focal Loss, dynamically amplifying the classification loss weights under high-speed operation, forcing the network to allocate more gradient resources for feature recognition, thereby significantly reducing the missed detection rate under high-speed conditions. Considering the significant "axially slender" geometric features of wire rope defects, and the fact that lateral (diameter direction) positioning accuracy is far more important than longitudinal (rope length direction) positioning accuracy, this method improves the distributed focus loss with anisotropic weighting, forcing the network to preferentially converge the boundary probability distribution along the wire rope diameter direction, ensuring that the detection box tightly wraps around the wire rope edge, and avoiding misidentification of background noise (such as shaft wall) as rope defects.

[0190] ;

[0191] ;

[0192] in, This represents the improved classification loss value; This indicates the number of positive samples (real defects) in the current batch, used for normalization; For sample index; The velocity sensitivity factor controls the degree to which the loss function is sensitive to velocity changes, and its value ranges from 0.5 to 1.0. The real-time running speed of the wire rope is fed back by the encoder; The rated maximum speed of the elevator; This is a class balancing parameter used to adjust the weights of positive and negative samples, addressing the problem of scarce wire rope defect samples; it is set to 0.25. This is a focusing parameter used to adjust the weights of easy and difficult samples. The larger the value, the more the network focuses on fuzzy defects that are difficult to classify; we take 2.0. Predict the probability that the sample belongs to the defect category for the network; This represents the improved regression distribution loss value; Indicates the focus weight in the width direction. This method sets the focusing weight in the length direction. This is to give higher priority to lateral positioning; Indicates the first The x and y coordinates of the center point of the actual bounding box; This represents the probability distribution of coordinates predicted by the network.

[0193] Based on the above analysis, the global optimization objective is calculated to achieve balanced regression, classification, and physical consistency tasks, guiding the network parameters to converge toward the global optimal solution. This represents the total loss function used for backpropagation; This represents the balance weighting factor for the losses of each component. .

[0194] To transform the deep features extracted from the backbone network and the physical neck into physical quantities usable for decision-making, the raw output of the decoupled detection head is first mathematically mapped. This involves extracting semantic features through independent convolutional branches, and then converting the obtained physical enhancement features... The mapping is defined in the continuous real space. Classification primitive tensor Among them, channel dimension This corresponds to a preset defect category probability distribution; simultaneously, it utilizes a regression branch to extract geometric location features, and combines this with a distribution focus loss to map the feature channels to a representation defined in the real number space. Distance distribution tensor in Its channel dimension 4n contains discrete probabilistic information used to achieve sub-pixel level positioning accuracy. The specific calculation formulas for the above feature decoupling and mapping are as follows:

[0195] ;

[0196] ;

[0197] For the classification feature branch, the unnormalized output is first processed using the Sigmoid activation function. Mapping to the (0,1) probability interval generates a standardized confidence level for subsequent multimodal fusion computation. ,in Representing infrared or ultrasonic modes, Indicates the first One target, the operation is as follows:

[0198] ;

[0199] Simultaneously, for the regression feature branch, in order to extract from the distribution features To accurately recover the spatial location of defects, this method employs an expectation integral strategy defined by the distributed focal loss. The regression distribution is transformed into a probability density function using the Softmax function, and the mathematical expectation is calculated. Subsequently, the discrete distance index is... The mathematical expectation is calculated by weighting the values ​​(ranging from 0 to n) to obtain the relative offset at the feature map scale. Furthermore, combining the downsampling step size S of the current feature map with the anchor point coordinates... The final predicted bounding box absolute coordinates are calculated. And the coordinate vector Includes the required center coordinates ( and size This forms the basis for subsequent loss calculations and decision-making:

[0200] , ;

[0201] The original confidence levels of the infrared and ultrasonic branches obtained from the aforementioned steps ( ) and bounding box coordinates ( The system then enters the physical consistency verification phase.

[0202] To address the common unilateral interference issues in elevator shaft environments, such as oil stain reflection (infrared artifacts) or mechanical vibration (ultrasonic noise), this method utilizes the homology of acoustic and thermal signals in physical space and constructs a cross-modal fusion confidence score using Formula 5. This computational model not only integrates the original confidence scores of a single mode, but also innovatively introduces the spatial intersection-over-union (IoU) ratio of the infrared and ultrasonic prediction frames as a strong physical verification factor. Its core criterion is that genuine structural damage is necessarily accompanied by both abrupt changes in acoustic impedance and localized frictional heating. Therefore, only when the prediction frames of the acoustic and thermal modes are highly overlapping spatially (i.e., a high IoU) will the fusion confidence score approach 1. Conversely, if the spatial intersection-over-union ratio is low, the detected signal is determined to be a non-homogeneous spurious artifact, and the system will automatically eliminate such false alarms through mathematical suppression, thereby ensuring the physical authenticity of the detection results. The fusion confidence score is calculated using Formula 5, which is:

[0203] ;

[0204] In the formula, Indicates the first The fusion confidence of the objectives; This represents the raw confidence level (0~1) of the infrared and ultrasonic branches output by the network. The spatial intersection-union ratio (IUU) of the infrared defect prediction box and the ultrasonic defect prediction box is based on the real-time decoding output. The absolute coordinate vector of the defect prediction bounding box for each target and Dynamically calculated geometric verification variables. The spatial intersection-union ratio is obtained using Formula 6. :

[0205] ;

[0206] In the formula, Indicates the area to be calculated. and Representing the corresponding number The spatial overlap area and total coverage area of ​​the acoustic and thermal two-mode prediction boxes for each target.

[0207] Based on the confirmation of the damage's authenticity, a comprehensive risk assessment model based on operational condition perception was constructed to further quantify the actual threat posed by the fault to elevator operational safety. According to the principles of fracture mechanics and tribothermodynamics, under high-speed operation, even minute defects of the same geometric size in a wire rope can often trigger severe thermal failure or crack propagation. Therefore, this method calculates the comprehensive risk index of the current detection frame using Formula 7. This formula, while accumulating the percentage of defect geometric area, introduces an exponentially increasing rate of risk amplification factor. This mathematical design enables dynamic risk assessment grading: in low-speed maintenance mode, the system maintains a certain tolerance for minor wear; while in high-speed operation mode, this factor significantly amplifies the risk weight of minor defects, thereby giving the system higher safety sensitivity under harsh operating conditions. In other words:

[0208] ;

[0209] In the formula, The pixel area of ​​the defect box; This represents the total effective area of ​​the wire rope within the current field of view; As a speed risk amplification factor, among which As a risk factor, the faster the speed, the greater the potential frictional heat hazard caused by a defect of the same size, and the higher the risk score.

[0210] Finally, based on the fusion confidence level and cumulative risk index, a tiered control instruction is output. :

[0211] .

[0212] This indicates a Level 1 shutdown command, which is only triggered when the infrared detection frame and the ultrasonic detection frame are highly overlapped in spatial position, and both have high confidence levels. Only then can the high threshold be broken. This effectively avoids false alarms from a single sensor, ensuring that only confirmed structural damage triggers an emergency shutdown. This indicates a Level 2 warning order. If the confidence level of a single point does not meet the shutdown standard, but the cumulative risk index after speed weighting exceeds the cumulative risk threshold, the warning is issued. If the condition is detected as intensive fatigue wear or corrosion, a level 2 early warning command will be output to prompt maintenance. This indicates normal operating status. In practical applications, the first-level shutdown threshold... The recommended value range is 0.85~0.95 to ensure extremely high confidence; cumulative risk threshold. It needs to be calibrated according to the wire rope diameter and the inspection cycle. A typical value can be set to 10.0.

[0213] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A real-time acoustic-thermal composite detection method for elevator wire ropes, characterized in that, The detection method includes the following steps: Multimodal data acquisition steps: Acquire ultrasonic signals inside the wire rope, infrared two-dimensional temperature data on the surface of the wire rope, and pulse code signals of the wire rope operation; The steps for constructing heterogeneous tensors include: The ultrasonic signal is subjected to a complex Morlet wavelet continuous transform to reconstruct a two-dimensional ultrasonic time-frequency matrix, and then extended into a high-dimensional ultrasonic time-frequency tensor. The infrared two-dimensional temperature data is subjected to adaptive statistical normalization processing and reconstructed into an infrared feature tensor; The pulse-coded signal is solved and normalized, and a physical manifold control vector is generated through nonlinear mapping; The constructed physical manifold sensing and dual-stream lightweight reassembly network is used to enter the multimodal data processing step, including: Ultrasonic features and infrared features are extracted from the ultrasonic time-frequency tensor and the infrared feature tensor respectively through a dual-stream backbone network, and then spliced ​​to generate a multimodal fusion feature tensor. In the feature fusion stage, a physical gain factor is generated based on the physical manifold control vector and fused with the multimodal fusion feature tensor to output the fused feature enhanced by physical perception. Based on the fusion features enhanced by physical sensing, the detection head outputs infrared defect prediction boxes and ultrasonic defect prediction boxes respectively. Decision-making and hierarchical management steps: Execute bimodal mutual verification logic to calculate the spatial intersection-union ratio (CIU) of the infrared defect prediction box and the ultrasonic defect prediction box; obtain the fusion confidence of the wire rope damage based on the spatial CIU. The comprehensive risk index of wire rope damage is calculated based on the introduced speed risk amplification factor; Based on the comprehensive risk index and fusion confidence level, the damage risk level is determined, and a graded control instruction corresponding to the risk level is output.

2. The real-time acoustic-thermal composite detection method for elevator wire ropes according to claim 1, characterized in that, The operation for constructing the ultrasonic time-frequency tensor includes: First, define a complex Morlet mother wavelet basis function. : ； in, The time independent variable of the basis function; Indicates bandwidth parameter; Indicates the center frequency; Represents the imaginary unit; Pi is a constant. Denotes the phase quadratic term factor. This represents the acoustic wave dispersion phase compensation coefficient; Construct physical frequency With mathematical scale The inverse mapping relationship between them: in, This indicates the sampling rate of the data acquisition card, which is the number of discrete data points collected per second. This represents the high-frequency attenuation adaptive gain term; Using the above basis functions The original ultrasound signal is convolved, and the correlation strength of the signal at different time points and frequency scales is extracted using Formula 1. Formula 1 is as follows: in, These are complex wavelet coefficients, representing the signal at a specific time. and specific scale Similarity; Represents the original ultrasound signal; This represents the translation parameter, corresponding to the time axis coordinates of the output time-frequency plot; Represents the mother wavelet function; The ultrasonic time-frequency tensor is calculated using Formula 2, which is: in, This represents the ultrasonic time-frequency tensor input to the network; This represents the modulo-square operation; This is the noise floor cutoff threshold; This is the nonlinear amplification factor; To filter out invalid background signals below the ambient noise level, This represents a numerical stability constant to prevent computational overflow caused by taking the logarithm of 0.

3. The real-time acoustic-thermal composite detection method for elevator wire ropes according to claim 1 or 2, characterized in that, The operations for constructing infrared feature tensors include: Define a nonlinear mapping model based on the statistical moments of a local sliding window: in, This represents the texture-enhanced infrared feature tensor that is ultimately input into the network; This represents the raw two-dimensional infrared temperature data; and Represented by pixels Centered The mean background temperature and the standard deviation of the temperature texture within a local neighborhood window; This represents a numerical stability constant to prevent the denominator from being zero. This represents the contrast enhancement factor.

4. The real-time acoustic-thermal composite detection method for elevator wire ropes according to claim 1 or 2, characterized in that, The operations for constructing physical manifold control vectors include: The pulse-coded signal includes real-time instantaneous speed and running distance. First, the real-time speed and running distance are normalized and periodically encoded: ； ； in, This represents the normalized velocity scalar; This indicates the real-time instantaneous speed of the encoder feedback; This indicates the elevator's rated maximum speed; Indicates the speed sensitivity coefficient; This represents the hyperbolic tangent activation function, which maps the velocity to the interval (-1, 1). Indicates the absolute mileage of the wire rope; Indicates the lay distance of the wire rope; Represents a positional periodic encoding vector; An initial physical feature vector containing velocity and phase is constructed using Formula 3, and a nonlinear manifold mapping is performed. Formula 3 is as follows: ； ； in, Represents the physical manifold control vector; This is the initial physical feature vector; This represents the learnable weight matrix of the multilayer perceptron. is the bias vector of the multilayer perceptron; This indicates a modified linear unit activation function that introduces nonlinear features.

5. The real-time acoustic-thermal composite detection method for elevator wire ropes according to claim 1, characterized in that, The operations for extracting both ultrasound and infrared features include: The infrared feature tensor and the ultrasonic time-frequency tensor are jointly defined as the input tensor. , Represents the infrared feature tensor. Represents the time-frequency tensor of ultrasound; Input tensor Divide into active and passive parts: ； ； in, Indicates the number of channels; This indicates the number of active channels involved in the calculation; The component involved in the calculation represents the feature component that has been segmented and is used in the actual spatial convolution calculation. This represents the remaining feature components that do not participate in the convolution operation and are directly subjected to identity mapping; Formula 4 is used to pair the eigencomponents of the active part. For feature extraction, Formula 4 is as follows: ； ； in, These are the extracted active space features; Output feature map; for Convolution kernel weight matrix; For texture direction prior operators; This represents the bias term of the convolutional layer; It is a learnable channel scaling factor used to adaptively adjust the feature amplitude to enhance weak defect signals.

6. The real-time acoustic-thermal composite detection method for elevator wire ropes according to claim 5, characterized in that, The operations for generating multimodal fusion feature tensors include: Introducing perceptual weights The infrared / ultrasound multi-scale aggregated feature tensor is output using a pyramid pooling module for weighted fusion. The calculation formula is as follows: ； ； In the formula, For infrared / ultrasound multiscale aggregated feature tensor; For the first Hierarchical pooling characteristics; It is the max pooling operator; Represents the convolution kernel; The values ​​are 1, 2, and 3. This indicates a weighted concatenation based on the channel dimension; The scale importance coefficients are automatically learned by the network during backpropagation; The output infrared multiscale aggregated feature tensor and ultrasonic multiscale aggregate feature tensor The data is concatenated along the channel dimension to generate a multimodal fusion feature tensor. .

7. The real-time acoustic-thermal composite detection method for elevator wire ropes according to claim 1 or 6, characterized in that, The operations that output the fused features enhanced by physical perception include: Employing a large convolutional kernel decomposition strategy to fuse feature tensors from multiple modalities Extract long-range spatial dependency features, namely: ； ； in, This is an intermediate value representing the horizontal characteristic. This represents the final long-range spatial dependency feature; This represents depthwise separable convolution; Indicates the size of the large convolution kernel; Represents asymmetric convolution kernels in the horizontal and vertical directions; The physical gain factor is generated based on the physical manifold control vector, and the operation is as follows: ； in, Indicates a velocity-dependent offset; Represents the physical manifold control vector; Indicates the weights and biases of the gated projection layer; Physical gain factor; Based on the above physical gain factor We construct horizontal attention and velocity-modulated vertical attention separately, thereby outputting fused features enhanced by physical perception: ； ； in, This indicates a fusion feature enhanced by physical perception. Indicates the use of channel information fusion convolution; The temperature coefficient for attention sharpening is introduced; Sigmoid represents the activation function. Indicates horizontal attention weights; This represents the vertical attention weight.

8. The real-time acoustic-thermal composite detection method for elevator wire ropes according to claim 1 or 2, characterized in that, The fusion confidence level of wire rope damage is obtained using Formula 5, which is: ； In the formula, Indicates the first The fusion confidence of the objectives; This represents the raw confidence level of the infrared and ultrasonic branches output by the network; The spatial intersection-union ratio (IUU) of the infrared defect prediction box and the ultrasonic defect prediction box is based on the real-time decoding output. The absolute coordinate vector of the defect prediction bounding box for each target and Dynamically calculated geometric verification variables; The spatial intersection-union ratio is obtained using Formula 6. : ； In the formula, Indicates the area to be calculated. and Representing the corresponding number The spatial overlap area and total coverage area of ​​the acoustic and thermal two-mode prediction boxes for each target.

9. The real-time acoustic-thermal composite detection method for elevator wire ropes according to claim 1 or 2, characterized in that, The construction of the physical manifold sensing and dual-stream lightweight reconfiguration network also includes obtaining the total loss function through Formula 8, which is: ； in, Represents the total loss function; The weighting factor representing the balance of losses in each component; Indicates the crossover and union ratio loss; Represents the SIoU loss function; This represents the improved classification loss value; This represents the improved regression distribution loss; This represents the final physical consistency loss value.

10. A real-time acoustic-thermal composite detection device for elevator wire ropes, characterized in that, The detection device includes: An ultrasonic transducer (1) is used to inject ultrasonic waves into a steel wire rope and receive the reflected ultrasonic signals. Infrared thermal imager (2) is used to collect infrared two-dimensional temperature data of the surface of the wire rope; A rotary encoder (3) is used to acquire pulse-coded signals of the wire rope in real time. The data processing terminal (4) is used to perform real-time tensor fusion and decision-making on ultrasonic signals, infrared two-dimensional temperature data and pulse code signals based on the physical manifold perception and dual-stream lightweight recombination network algorithm integrated within it, so as to realize the identification and graded alarm of wire rope damage. The ultrasonic transducer (1), infrared thermal imager (2), and rotary encoder (3) are all connected to the data processing terminal (4).

Images