Fault detection method and related device

Abstract

The present disclosure provides a fault detection method and related equipment, which belongs to the technical field of computers. The method comprises: obtaining vibration signals of a to-be-detected object in at least two directions; obtaining initial health detection results of the to-be-detected object in at least two directions according to the vibration signals of the to-be-detected object in at least two directions; and fusing the initial health detection results of the to-be-detected object in at least two directions to determine a health detection result of the to-be-detected object. The scheme provided in the present disclosure can improve the efficiency and accuracy of quality detection and save labor costs.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and more specifically, to a fault detection method, apparatus and system, electronic device, computer-readable storage medium and computer program product. Background Technology

[0002] In recent years, with the gradual expansion of cities, more and more infrastructure, such as power cables, natural gas pipelines, and sewage systems, has been placed underground. To facilitate the maintenance of these facilities, roads are often equipped with numerous covered structures. Due to the prolonged pressure from passing vehicles, these covered structures and their underlying supporting structures are easily damaged. If they are not repaired and replaced in a timely manner, they will threaten road safety. Therefore, early inspection of the health condition of these covered structures is of great importance in mitigating potential risks.

[0003] In related technologies, the health status inspection of multiple covers mainly adopts a periodic inspection method. This operation, which requires frequent opening of multiple covers, not only increases the burden on maintenance personnel and maintenance costs, but also causes unnecessary interference to normal traffic on the road. At the same time, this method is inefficient, the inspection accuracy heavily depends on the experience of maintenance personnel, and it is easy to overlook hidden damage that cannot be visually detected, making it unsuitable for large-scale health status inspections of multiple covers. Summary of the Invention

[0004] The fault detection methods, apparatus and systems, electronic devices, computer-readable storage media and computer program products provided in this disclosure are intended to improve the efficiency and accuracy of fault detection and reduce costs.

[0005] This disclosure provides a fault detection method, including: acquiring vibration signals of an object to be detected in at least two directions; obtaining initial health detection results of the object to be detected in at least two directions based on the vibration signals of the object to be detected in at least two directions; and fusing the initial health detection results of the object to be detected in at least two directions to determine the health detection result of the object to be detected.

[0006] This disclosure provides a fault detection device, comprising: a receiving unit for acquiring vibration signals of an object to be detected in at least two directions; a processing unit for obtaining initial health detection results of the object to be detected in at least two directions based on the vibration signals of the object to be detected in at least two directions; the processing unit is further configured to fuse the initial health detection results of the object to be detected in at least two directions to determine the health detection result of the object to be detected.

[0007] This disclosure provides a fault detection system, including: a vibration sensor, an object to be detected being within the coverage area of ​​the vibration sensor, and a distance between the object to be detected and the vibration sensor, the vibration sensor being used to detect vibration signals of the object to be detected in at least two directions; and a fault detection device as described in any embodiment of this disclosure, wherein the receiving unit is used to acquire the vibration signals of the object to be detected detected by the vibration sensor in at least two directions.

[0008] This disclosure provides an electronic device, including: one or more processors; and a memory configured to store one or more programs, which, when executed by the one or more processors, cause the electronic device to implement the method described in any embodiment of this disclosure.

[0009] This disclosure provides a computer-readable storage medium having a computer program stored thereon, which, when run on a computer, causes the computer to execute the methods described in any embodiment of this disclosure.

[0010] This disclosure provides a computer program product, including a computer program that, when executed by a computer, implements the methods described in this disclosure.

[0011] The fault detection method, apparatus, system, electronic device, computer-readable storage medium, and computer program product provided in this disclosure, on the one hand, determine the health detection result of the object under test by acquiring vibration signals of the object under test in at least two directions, without having to frequently open the object under test, thus improving the efficiency of fault detection and reducing labor and time costs; on the other hand, by first obtaining the initial health detection result of the object under test in at least two directions based on the vibration signals of the object under test in at least two directions, and then fusing the initial health detection results in at least two directions to determine the health detection result of the object under test, the final health detection result does not rely on human experience, nor does it directly depend on the vibration signal in a single direction, thereby improving the accuracy of fault detection. Attached Figure Description

[0012] Figure 1 This is a flowchart of a fault detection method according to an exemplary embodiment of the present disclosure.

[0013] Figure 2 This is a schematic diagram of a time-frequency image input to a fault identification model according to an exemplary embodiment of the present disclosure.

[0014] Figure 3This is a schematic diagram illustrating the change in accuracy of a fault identification model as it undergoes training iterations, according to an exemplary embodiment of this disclosure.

[0015] Figure 4 This is a schematic diagram illustrating the change of the loss function of a fault identification model as it undergoes training iterations, according to an exemplary embodiment of this disclosure.

[0016] Figure 5 This is a schematic diagram of the network structure of a fault identification model according to an exemplary embodiment of this disclosure.

[0017] Figure 6 This is a schematic diagram of a visual interface of an exemplary embodiment of this disclosure.

[0018] Figure 7 This is a flowchart of a fault detection method according to another exemplary embodiment of this disclosure.

[0019] Figure 8 This is a schematic diagram of the structure of a fault detection device according to an exemplary embodiment of the present disclosure.

[0020] Figure 9 This is a schematic diagram of the structure of a fault detection system according to an exemplary embodiment of the present disclosure.

[0021] Figure 10 The illustration shows an application scenario of a fault detection system according to an embodiment of the present disclosure. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this disclosure more apparent, exemplary embodiments according to this disclosure will now be described in detail with reference to the accompanying drawings. In the drawings, the same reference numerals denote the same elements throughout. It should be understood that the embodiments described herein are merely illustrative and should not be construed as limiting the scope of this disclosure.

[0023] In this disclosure, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0024] In this disclosure, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance; the term "multiple" refers to two or more unless otherwise expressly defined. The terms "install," "connect," "link," and "fix" should be interpreted broadly. For example, "connect" can be a fixed connection, a detachable connection, or an integral connection; "link" can be a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in this disclosure according to the specific circumstances.

[0025] In the description of this specification, the terms "one embodiment," "some embodiments," "specific embodiment," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0026] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Various modifications and variations can be made to this disclosure by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

[0027] Figure 1 This is a flowchart of a fault detection method according to an exemplary embodiment of the present disclosure. Figure 1 The fault detection method provided in the embodiments can be executed by any electronic device, such as a terminal and / or a server, and this disclosure does not limit it. Figure 1 As shown, the method provided in this embodiment includes the following steps.

[0028] In S110, vibration signals of the object to be detected in at least two directions are acquired.

[0029] The object to be detected in this embodiment can be any object that needs to be detected for faults or defects. In an exemplary embodiment, the object to be detected includes multiple covers. Although multiple covers are used as examples in the following embodiments, this disclosure is not limited thereto.

[0030] In this disclosure, "multi-part cover(s) (MPC)" refers to a combined structure with multiple covers used to cover and protect internal facilities or internal spaces, or to allow maintenance personnel to enter the internal space for installation, maintenance, inspection, or other work, or to prevent accidental falls of personnel or objects. Examples include multi-part manhole covers, multi-part covers in drainage facilities, and multi-part covers covering the inlet and outlet of manholes. The specific design and function of the multi-part cover can vary depending on the application scenario and requirements. In the following embodiments, examples are provided of multi-part covers deployed on the road surface to cover and protect infrastructure (such as power cables, natural gas pipelines, sewage systems, etc.) under the road to facilitate maintenance of the infrastructure under the road, but this disclosure is not limited to this.

[0031] In an exemplary embodiment, acquiring vibration signals of the object to be detected in at least two directions includes: acquiring vibration signals in at least two directions collected by vibration sensors when a vehicle passes over the multi-cover surface.

[0032] In this embodiment of the disclosure, when the multi-cover surface is deployed on a road, various vehicles may drive over it. These vehicles may include, for example, cars, trucks, motorcycles, bicycles, electric vehicles, and other devices or equipment with transportation functions.

[0033] In this embodiment of the disclosure, vibration sensors can be deployed to collect vibration signals from the multi-coverage surface in multiple directions. Here, a vibration sensor refers to a sensor capable of detecting vibration signals of the object under test. The detected vibration signal refers to a signal including amplitude, phase, and frequency sensed when the object under test vibrates; that is, the vibration signal is a record of the changes in physical quantities (such as amplitude, phase, and frequency) generated when the object under test vibrates over time. In some embodiments, the vibration signal refers to the signal detected by the vibration generated when a vehicle passes over the multi-coverage surface. This can prevent interference from other environmental noise or ambient noise.

[0034] In the following embodiments, an example is given of using a triaxial accelerometer to simultaneously detect vibration signals of an object under test, such as a multi-faceted object, in the X / Y / Z directions. The triaxial accelerometer can simultaneously detect vibration signals of the object under test in three directions, providing comprehensive vibration data. However, this disclosure is not limited to this; for example, vibration signals of the object under test in two directions, or vibration directions in four or more directions, can be collected.

[0035] In three-dimensional space or a three-dimensional coordinate system, the X-axis typically represents the horizontal or forward / backward direction. In vibration detection, vibration signals along the X-axis can reflect the motion of the object being tested on a horizontal plane. The Y-axis typically represents the vertical direction (in some cases, it also represents the left-right direction, but here we will primarily explain the vertical direction). Vibration signals along the Y-axis can reflect the motion of the object being tested in the vertical direction, such as bouncing up and down or swaying left and right. The Z-axis typically represents another vertical direction (perpendicular to the Y-axis). In vibration detection, vibration signals along the Z-axis provide information about the motion of the object being tested in another vertical direction. However, in certain specific applications, the Z-axis can specifically refer to vertical vibrations above and below the ground surface.

[0036] In an exemplary embodiment, acquiring vibration signals in at least two directions collected by a vibration sensor when a vehicle passes over the multi-cover surface includes: acquiring the vibration signals in at least two directions collected by the vibration sensor upon receiving a trigger signal from a photoelectric sensor. The trigger signal from the photoelectric sensor indicates that a vehicle has passed over the multi-cover surface.

[0037] In this embodiment, automatic triggering of vibration signal acquisition can be achieved by deploying photoelectric sensors. In one scenario, the vibration sensor can acquire the vibration signal of the object under test in real time. When it receives a trigger signal from the photoelectric sensor, i.e., when a vehicle is detected passing over multiple surfaces, it sends out the vibration signal acquired at that moment, for example, to an oscilloscope for display and subsequent processing, i.e., fault detection. In another scenario, a trigger switch can be set on the vibration sensor. This switch is only activated when a trigger signal from the photoelectric sensor is received, and the vibration sensor begins detecting vibration signals from multiple surfaces. This reduces the power consumption of the vibration sensor and extends its service life. Simultaneously, automatic vibration signal detection can be achieved, eliminating the need for manual activation of the vibration sensor. Furthermore, this method ensures that the acquired vibration signal is generated by vibrations caused by vehicles passing over multiple surfaces, reducing the impact of environmental noise.

[0038] In an exemplary embodiment, the multiple covers are disposed on the road, and the vibration sensor is deployed on the sidewalk, with the multiple covers within the coverage area of ​​the vibration sensor.

[0039] In some embodiments, multiple covers are positioned on the road, and vibration sensors are deployed on the sidewalk (which can be any location on the sidewalk that does not affect normal pedestrian traffic). The vibration sensors can effectively detect the vibration signals of the multiple covers, meaning the multiple covers are within the coverage area of ​​the vibration sensors. There is a certain distance between the vibration sensors and the multiple covers, but this distance is not large enough to exceed the effective detection range of the vibration sensors. This achieves "short-distance remote sensing," that is, the vibration signals of the multiple covers are detected by the vibration sensors at a relatively close distance. On the one hand, the relatively close distance reduces interference; on the other hand, the detection process does not require direct contact with the object being detected, such as the multiple covers, preventing interference with traffic on the road and avoiding damage to the multiple covers. In other embodiments, the multiple covers can be positioned at any suitable location, and the vibration sensors can be deployed at any suitable location at a certain distance from them.

[0040] In this embodiment, the distance between the multi-coverage surface and the vibration sensor can be determined according to the specific usage. For example, the distance between the multi-coverage surface and the vibration sensor can be any distance not exceeding 14.6 meters, specifically based on the distance between the sidewalk and the multi-coverage surface, with the maximum distance being (3.65 × 4) meters for a four-lane elevated highway. A vibration sensor with an appropriate range, frequency response range, and sensitivity is selected according to the measurement requirements. The range of the vibration sensor covers the expected vibration range, the frequency response range should match the frequency characteristics of the vibration, and the sensitivity determines the measurement accuracy. The vibration sensor is placed at a suitable location along the roadside to ensure coverage of the multi-coverage surface to be detected.

[0041] In this embodiment, the vibration sensor is securely and reliably installed to minimize interference or damage from external factors. For example, an aluminum plate is first fixed to the ground using epoxy resin adhesive on the site (such as a selected location on a sidewalk). The aluminum plate serves as the mounting point for the triaxial accelerometer. Detection is performed after the epoxy resin has cured. A photoelectric sensor can also be deployed at a suitable location on the sidewalk to detect whether vehicles are passing over the multi-cover surface. This disclosure does not limit the installation location of the photoelectric sensor, as long as it does not obstruct normal vehicle and pedestrian traffic and can detect whether vehicles are passing over the multi-cover surface. Whenever a vehicle passes a designated location (above the multi-cover surface), the photoelectric sensor is triggered and instructs a data acquisition oscilloscope to record the vibration signal collected by the vibration sensor from the multi-cover surface.

[0042] On busy streets, ambient noise is unavoidable. In this embodiment, background noise is detected as a control group at the start of each experiment, and then wavelet denoising is used to filter out the noise. In addition to serving as a semi-automatic / automatic trigger condition, the photoelectric sensor also ensures that the collected vibration signals are emitted when vehicles pass over the surface.

[0043] In S120, based on the vibration signals of the object to be tested in at least two directions, the initial health detection results of the object to be tested in at least two directions are obtained.

[0044] In this embodiment of the disclosure, the initial health detection result of the object under test in each detected direction can be obtained first. For example, based on the vibration signal in the X direction, the initial health detection result of the multiple covers in the X direction is obtained, that is, the presence of a fault in the multiple covers is determined based on the vibration signal in the X direction. Based on the vibration signal in the Y direction, the initial health detection result of the multiple covers in the Y direction is obtained, that is, the presence of a fault in the multiple covers is determined based on the vibration signal in the Y direction. Based on the vibration signal in the Z direction, the initial health detection result of the multiple covers in the Z direction is obtained, that is, the presence of a fault in the multiple covers is determined based on the vibration signal in the Z direction.

[0045] In S130, the initial health detection results of the object to be detected in at least two directions are fused to determine the health detection result of the object to be detected.

[0046] In this embodiment of the disclosure, the initial health detection results of the object to be detected in multiple directions are considered simultaneously to obtain the final health detection result of the object to be detected, that is, whether the object to be detected has a fault or does not have a fault.

[0047] In an exemplary embodiment, fusing the initial health detection results of the object to be detected in at least two directions to determine the health detection result of the object to be detected includes: taking the initial health detection results of the object to be detected in each direction as corresponding evidence bodies; calculating the similarity between evidence bodies; obtaining the mutual support between evidence bodies based on the similarity between evidence bodies; obtaining the weight coefficient of the evidence bodies based on the mutual support between evidence bodies; and determining the health detection result of the object to be detected based on the evidence bodies and their corresponding weight coefficients.

[0048] In the following embodiments, the use of Dempster-Shafer (DS) evidence theory to fuse initial health test results from multiple surfaces in the X, Y, and Z directions is illustrated, but this disclosure is not limited thereto. Embodiments of this disclosure utilize DS evidence theory to help integrate multiple diagnostic information (initial health test results of the subject in multiple directions), effectively merging information from multiple independent evidence bodies or sources, thereby providing a more accurate diagnostic result (the final health test result of the subject).

[0049] In traditional DS evidence theory, all evidence or evidentiary bodies are considered to have the same weight, meaning they are equally important in the fusion process. However, in practical applications, different pieces of evidence often differ in importance due to variations in source, reliability, precision, and other factors. Therefore, traditional DS evidence theory may produce unreasonable fusion results when dealing with evidence of varying importance, especially when there are conflicting characteristics between different focal points of evidence.

[0050] In this embodiment of the disclosure, the improved DS evidence theory is used to perform decision-level fusion of the diagnostic results (initial health test results of the multi-coverage surface in the X / Y / Z directions) in the three directions.

[0051] Specifically, firstly, the initial health test results of the multi-coverage surface in the X / Y / Z directions are used as corresponding evidence.

[0052] For example, based on the vibration signals of multiple covers in the X, Y, and Z directions, time-frequency images of the multiple covers in the X, Y, and Z directions are obtained. These time-frequency images are then input into a fault identification model. The fault identification model processes these images in the X, Y, and Z directions respectively, outputting p(X) (representing the probability of a fault in the multiple covers predicted based on the time-frequency image in the X direction) and 1-p(X) (representing the probability of no fault in the multiple covers predicted based on the time-frequency image in the X direction). If p(X) is greater than 1-p(X), then the initial health detection result determined based on the vibration signals in the X direction is that the multiple covers are faulty. If the cover surface is faulty, then the multiple cover surfaces are not faulty; p(Y) (representing the probability of the multiple cover surfaces being faulty based on the time-frequency image in the Y direction), 1-p(Y) (representing the probability of the multiple cover surfaces not being faulty based on the time-frequency image in the Y direction), if p(Y) is greater than 1-p(Y), then the initial health detection result based on the vibration signal in the Y direction indicates that the multiple cover surfaces are faulty, otherwise, the multiple cover surfaces are not faulty; p(Z) (representing the probability of the multiple cover surfaces being faulty based on the time-frequency image in the Z direction), 1-p(Z) (representing the probability of the multiple cover surfaces not being faulty based on the time-frequency image in the Z direction) are six probability values. If p(Z) is greater than 1-p(Z), then the initial health detection result based on the vibration signal in the Z direction indicates that the multiple cover surfaces are faulty, otherwise, the multiple cover surfaces are not faulty. p(X) and 1-p(X) can be used as the first evidence of vibration signals from the X direction; p(Y) and 1-p(Y) can be used as the second evidence of vibration signals from the Y direction; and p(Z) and 1-p(Z) can be used as the third evidence of vibration signals from the Z direction.

[0053] The evidence body is the Basic Probability Assignment (BPA), also known as the m-function, associated with propositions in the identification framework. It assigns a degree of confidence to each proposition, representing the level of confidence in that proposition. In DS evidence theory, the evidence body is embodied through the BPA, which represents the degree of confidence or support for a particular proposition (here referring to whether multiple facets are faulty or not).

[0054] Then, the similarity between any two pieces of evidence is obtained. For example, the similarity between the i-th piece of evidence can be calculated using the following formula. and the j-th piece of evidence Similarity between

[0055]

[0056] In formula (1) above, the values ​​of i and j depend on the number of evidence bodies. Assuming there are n evidence bodies in total, i.e., n is the number of independent evidence bodies, then i and j are both positive integers greater than or equal to 1 and less than or equal to n, where n is a positive integer greater than 1. Here, k represents the number of propositions corresponding to each evidence body, specifically the number of diagnostic result categories corresponding to each evidence body, and k is a positive integer greater than 1. The value of h is a positive integer greater than or equal to 1 and less than or equal to k. m ih m represents the degree of confidence in the h-th proposition within the i-th piece of evidence. jh This indicates the degree of confidence in the h-th proposition within the j-th piece of evidence.

[0057] For example, assuming three evidence bodies are used in the X / Y / Z directions of the multi-coverage surface, then n=3, meaning the number of evidence bodies corresponds to the number of vibration signals in each direction. k=2, meaning there are two propositions (assuming h=1 indicates a fault in the multi-coverage surface; h=2 indicates no fault in the multi-coverage surface), then m 11 m represents the probability that multiple surfaces in the first piece of evidence are faulty. 12 This represents the probability that there is no fault in the multiple covers of the first piece of evidence, and so on for the others. That is, there are 3 pieces of evidence, each corresponding to two probabilities. The first piece of evidence is: m 11 =p(X),m 12 =1-p(X); The second piece of evidence is: m 21 =p(Y),m 22 =1-p(Y); The third piece of evidence is: m 31 =p(Z),m 32 =1-p(Z).

[0058] It is understood that the above example is based on distinguishing only two classification results: the presence of faults on multiple surfaces and the absence of faults. In other embodiments, if the fault identification model can distinguish more classification results, the value of k will increase accordingly.

[0059] After calculating the similarity between any two pieces of evidence based on the above formula (1), the similarity matrix shown in the following formula (2) can be obtained:

[0060]

[0061] Then, the mutual support between the other evidence pieces (excluding the i-th evidence piece) and the i-th evidence piece can be obtained using the following formula (3).

[0062]

[0063] Then, using the following formula (4), the weight coefficient ω of the i-th piece of evidence is obtained based on the mutual support of the i-th piece of evidence. i :

[0064]

[0065] In this embodiment, the mutual support between pieces of evidence represents the degree to which one piece of evidence is supported by the other pieces of evidence. A higher support indicates a more important piece of evidence, and thus a greater weight is assigned to it; a lower support indicates a less important piece of evidence, and thus a smaller weight is assigned to it. Using a strategy of determining trust level based on mutual support, formula (3) can be understood as a formula for calculating the trust level of other pieces of evidence towards the i-th piece of evidence. A higher support indicates that other pieces of evidence trust that piece of evidence more.

[0066] It is understood that any weighting method based on trust level falls within the scope of this disclosure; that is, the higher the trust level, the larger the weighting coefficient assigned to the evidence. For example, the weighting coefficient ω of the i-th piece of evidence can also be determined using the following formula. i :

[0067]

[0068] Wherein, I in the above formula i and I j The importance of the i-th and j-th pieces of evidence can be determined through expert evaluation.

[0069] For example, the weight coefficient ω of the i-th piece of evidence can be determined based on the distance between different pieces of evidence. i :

[0070]

[0071] In the above formula, d ij This represents the distance between the i-th and j-th pieces of evidence. For example, it can be calculated using Euclidean distance, but this disclosure is not limited to this. i Let s represent the sum of similarities between the i-th piece of evidence and all other pieces of evidence except the i-th piece of evidence. ij This represents the similarity between the i-th piece of evidence and the j-th piece of evidence.

[0072] For example, weight coefficients can also be assigned based on the fuzzy weights of the evidence pieces. Let the fuzzy weight of the i-th piece of evidence be... The weight coefficient ω of the i-th piece of evidence can be represented using triangular fuzzy numbers. i :

[0073]

[0074] Finally, new evidence was obtained.

[0075]

[0076] In this embodiment, a fault identification model is first used to diagnose vibration signals in the X, Y, and Z directions, resulting in three independent evidence bodies. Then, the weight coefficient of each evidence body is calculated according to the above formula (4). The greater the support from the other two directions in a given direction, the greater the weight coefficient of the corresponding evidence body. Then, the three evidence bodies are weighted and summed based on formula (5) to obtain a new evidence body, that is, the fusion result is obtained using the improved DS evidence theory synthesis rule (whether it belongs to faulty or faultless, and which classification result has a higher probability).

[0077] For example, the above formula (5) can be expressed as:

[0078]

[0079] Formula (6) is the probability of a fault in the multi-cover surface predicted after fusing vibration signals from three directions. Formula (7) is the probability of no fault in the multi-cover surface predicted after fusing vibration signals from three directions. By comparison... and The size of both is used to determine the final health test result of the multi-coverage surface, that is, if Greater than If the final health test result indicates that the multi-cover surface is faulty, then the multi-cover surface is not faulty.

[0080] The method provided in this disclosure assigns different weight coefficients to different pieces of evidence based on their importance. During the evidence fusion process, more important and reliable evidence is given higher weight coefficients, thereby reducing the impact of unimportant or unreliable evidence on the fusion result (i.e., the final health test result). Introducing weight coefficients into the synthesis rules of the DS evidence theory ensures that the contribution of each piece of evidence is proportional to its weight coefficient during fusion. By introducing weight coefficient allocation, the improved DS evidence theory effectively avoids the weakening of feature conflicts between different evidence focuses. Because more important and reliable evidence is given higher weight coefficients, the fusion result better reflects the true situation, thus improving the accuracy of the fusion result. Even if some unreliable or erroneous evidence exists, its lower weight coefficients minimize its impact on the fusion result, thereby enhancing the robustness of the system. Furthermore, by adjusting the weight coefficient values ​​of the evidence, the system can flexibly adapt to the needs of different application scenarios, improving its adaptability.

[0081] In related technologies, some studies have attempted to use drones or satellites to photograph multiple covers and inspect their surfaces using visual methods. However, the image quality obtained using this method is easily affected by external factors such as weather, leading to misjudgments. Furthermore, this method fails to effectively detect defects located within the multiple covers. Other studies have used laser scanners to scan multiple covers to analyze their health status, but this method is costly and impacts road safety.

[0082] In an exemplary embodiment, obtaining initial health detection results of the object under test in at least two directions based on the vibration signals of the object under test in at least two directions includes: performing continuous wavelet transform on the vibration signals of the object under test in each direction to obtain time-frequency images of the object under test in each direction; and inputting the time-frequency images of the object under test in each direction into a fault identification model to obtain initial health detection results of the object under test in each direction.

[0083] For example, continuous wavelet transforms are performed on the vibration signals of the multi-cover surface in the X, Y, and Z directions to obtain wavelet coefficients in each direction, thereby constructing time-frequency images in each direction. These time-frequency images in the X, Y, and Z directions are then input into a trained fault identification model to determine whether the initial health detection result of the multi-cover surface in each X, Y, and Z direction indicates the presence or absence of a fault.

[0084] The fault detection model in this disclosure refers to a machine learning model, neural network model, or deep learning model that can identify whether a detected object has a fault based on an input time-frequency image. In the following embodiments, a CNN (Convolutional Neural Networks) model is used as an example, but this disclosure is not limited to this, and any other model capable of classification and recognition can be used.

[0085] This disclosure discloses an intelligent fault detection method developed using artificial intelligence technology. When applied to fault detection of multiple road surfaces, it can also be referred to as an intelligent multi-surface health status inspection method. This method can quickly and accurately understand the health status of multiple road surfaces. Furthermore, the short-distance data collection method at pedestrian walkways does not interfere with traffic, effectively reducing detection costs caused by traffic disruptions.

[0086] In an exemplary embodiment, performing continuous wavelet transforms on the vibration signals of the object to be detected in each direction to obtain time-frequency images of the object to be detected in each direction includes: determining a mother wavelet function including waveform adjustment parameters; determining a basis wavelet function based on the mother wavelet function including waveform adjustment parameters, wherein the waveform adjustment parameters are used to adjust the decay rate of the basis wavelet function; performing continuous wavelet transforms on the vibration signals of the object to be detected in each direction based on the basis wavelet function to obtain wavelet coefficients of the object to be detected in each direction; and determining the time-frequency images of the object to be detected in each direction based on the wavelet coefficients of the object to be detected in each direction.

[0087] In an exemplary embodiment, the mother wavelet function φ includes waveform adjustment parameters. α (t) is expressed by the following formula:

[0088]

[0089] Where μ is the undetermined normalization coefficient, taking values ​​in the range [0,1], and can be represented by the coefficient π in the Morlet wavelet function. 1 / 4 Take several values ​​and observe the artifact phenomenon of CWT (which tends to appear at the boundaries and produce boundary effects). If the artifact phenomenon is relatively serious, gradually increase or decrease the normalization coefficient, and select the normalization coefficient corresponding to the time-frequency image with better effect (time-frequency image with relatively less serious artifact phenomenon) as the final normalization coefficient; α is the waveform adjustment parameter; ξ is the damping ratio parameter, ξ is a real number greater than 0 and less than 1, which can be obtained based on experience and selected in conjunction with the normalization coefficient; t represents time; ω represents angular frequency.

[0090] In practical applications, vibration signals are finite, not infinitely long. When wavelet transform is performed at the boundaries of the vibration signal, the lack of sufficient data support leads to boundary effects. Furthermore, in continuous wavelet transforms, a fixed-size window is typically used. When the window moves to the boundary of the vibration signal, part of the window may extend beyond the signal's range, resulting in boundary effects. Boundary effects in continuous wavelet transforms refer to the discontinuities or distortions that occur at the boundaries of the vibration signal due to the finite length of the vibration signal and the finite size of the transform window. Specifically, when the wavelet function is applied at the boundary of a signal, insufficient data at the signal's end to support the complete wavelet transform results in boundary effects.

[0091] In an exemplary embodiment, taking the Morlet wavelet function as the mother wavelet function as an example, its expression is:

[0092]

[0093] In this embodiment, the Morlet wavelet function is improved to become the above formulas (8) and (9), that is, the improved mother wavelet function is φ. α (t). It is understood that the specific form of the mother wavelet function is not limited in the embodiments of this disclosure, and any suitable mother wavelet function can be adopted. The mother wavelet function is an oscillating function with zero mean and fast decay, which satisfies integrability and square integrability.

[0094] In some embodiments, the improved mother wavelet function η1(t) can also be used as shown below:

[0095] η1(t)=0.6*φ(t)+0.4*θ1(t) (10-2)

[0096]

[0097] In other embodiments, an improved mother wavelet function, as shown below, may also be used.

[0098] η2(t):

[0099] η2(t)=0.6*φ(t)+0.4*θ2(t) (10-4)

[0100]

[0101] Among them, the damping ratio parameter mainly affects the decay rate and shape of the wavelet. For transient events or oscillations, a high damping ratio parameter can be used to detect sharp transients or transient features in signals, and is more suitable for use in the embodiments of this disclosure. It can effectively reduce the boundary effect in continuous wavelet transform and improve the accuracy and reliability of the transform.

[0102] In this embodiment of the disclosure, the undetermined normalization coefficient is used for normalization. Since the peak value of each collected vibration signal is different, normalization is required when processing the waveform to maintain the consistency of total energy (i.e., energy conservation) to ensure comparability when analyzing different parts of different vibration signals or multiple different vibration signals.

[0103] The mother wavelet function used in this embodiment is an improved version of the traditional Morlet wavelet function. α (t), and then according to the mother wavelet function φ α (t) Determine the basis wavelet function. Then, integrate the vibration signals f(X), f(Y), and f(Z) in the three directions with this basis wavelet function to obtain the wavelet coefficients for each of the three directions.

[0104] For example, taking the vibration signal f(X) as an example, its continuous wavelet transform can be expressed by the following formula:

[0105]

[0106] In the above formula, Here, is the basis wavelet function, 'a' is the scaling parameter, which can be used to compress or stretch the basis wavelet function, thus detecting different frequency components in the vibration signal. 'a' is greater than 0. 'b' is the translation parameter, which is used to move the basis wavelet function on the time axis, allowing analysis of the vibration signal's behavior at different time points. yes The conjugate; These are wavelet transform coefficients, or simply wavelet coefficients.

[0107] In this embodiment of the disclosure, when the decay rate of the basis wavelet function is adjusted by waveform modulation, especially by increasing the decay rate, the overlap between the basis wavelet function and the boundary of the vibration signal can be effectively reduced. On the one hand, since the extension of the basis wavelet function outside the boundary is reduced, the vibration signal at the boundary will not be subjected to excessive external interference, thereby reducing artifacts. On the other hand, a basis wavelet function with faster decay can more accurately resolve the features at the boundary of the vibration signal, so that the transformation result still has high accuracy in the edge region of the vibration signal.

[0108] In an exemplary embodiment, performing continuous wavelet transform on the vibration signals of the object to be detected in each direction to obtain time-frequency images of the object to be detected in each direction includes: performing noise filtering on the vibration signals of the object to be detected in each direction to obtain the vibration signals of the object to be detected after noise removal in each direction; and performing continuous wavelet transform on the vibration signals of the object to be detected after noise removal in each direction to obtain time-frequency images of the object to be detected in each direction.

[0109] Figure 2 Examples are shown of some 2D (2-dimensional) time-frequency images obtained after performing continuous wavelet transform, i.e., CWT transformation, which can be input into a fault identification model for processing. Figure 2 The horizontal axis represents time, and the vertical axis represents frequency. Figure 2 The distribution of wavelet coefficients in the time and frequency domains is shown.

[0110] In an exemplary embodiment, noise filtering is performed on the vibration signals of the object to be detected in each direction to obtain the noise-removed vibration signals of the object to be detected in each direction. This includes: decomposing the vibration signals of the object to be detected in each direction using a three-layer adaptive wavelet packet decomposition method to obtain wavelet coefficients; calculating the local variance of each wavelet coefficient; determining the noise intensity of each wavelet coefficient based on the local variance of each wavelet coefficient; determining the noise filtering threshold of the corresponding wavelet coefficient based on the noise intensity of each wavelet coefficient; and reconstructing wavelet coefficients with values ​​greater than the corresponding noise filtering threshold to obtain the noise-removed vibration signals of the object to be detected in each direction.

[0111] In some embodiments, the acquired vibration signals of the multi-covered surface in the X, Y, and Z directions are first denoised. For example, a three-layer adaptive wavelet packet decomposition method is first used to decompose the time-domain signals (i.e., vibration signals) in the X, Y, and Z directions; secondly, the local variance of each wavelet coefficient is calculated to estimate the noise intensity, and then an appropriate noise filtering threshold is selected; finally, wavelet coefficients with values ​​greater than the noise filtering threshold are retained for reconstruction to obtain clean vibration signals in the X, Y, and Z directions. Then, continuous wavelet transform is performed on the clean vibration signals in the X, Y, and Z directions to obtain time-frequency images in the X, Y, and Z directions.

[0112] Wavelet packet decomposition can decompose vibration signals in various directions into a series of wavelet coefficients in different sub-bands. The wavelet coefficients in different sub-bands represent information from different frequencies, and the wavelet coefficients / parameters in each sub-band also represent information from different time periods. In wavelet packet decomposition, vibration signals in various directions are typically decomposed into relatively large wavelet coefficients, while noise manifests as wavelet coefficients that are uniformly distributed across different sub-bands and have smaller values. Therefore, by processing the vibration signal and decomposing the wavelet coefficients to remove noise, the main characteristics of the vibration signal can be preserved.

[0113] Wavelet transform only decomposes the approximate parts of each level, while leaving the details unchanged. Therefore, the resolution of wavelet transform is not high enough. Wavelet packet decomposition, on the other hand, improves the time-frequency localization analysis capability of a signal by performing a more refined decomposition of the high-frequency components (noise signals are often high-frequency signals). Essentially, wavelet packet decomposition involves passing the original vibration signal through a high-pass filter and a low-pass filter to obtain high-frequency and low-frequency components, respectively. This process is repeated until a suitable frequency accuracy range is achieved. Wavelet packet decomposition decomposes the original vibration signal into 2^m (where m is a positive integer greater than or equal to 1) frequency bands at different levels. That is, wavelet packet decomposition is essentially a filter; by setting a noise filtering threshold, it filters out the wavelet coefficients of signals with small amplitudes, obtaining the wavelet coefficients of a clean signal. An inverse transform then yields a clean vibration signal. Finally, a continuous wavelet transform is applied to the clean vibration signal to obtain a two-dimensional CWT (Continuous Wavelet Transform) time-frequency image.

[0114] Three-level adaptive wavelet packet decomposition may include the following steps: Step 1. Select a basis wavelet function, such as the Daubechies wavelet function. This basis wavelet function may be the same as or different from the basis wavelet function of the continuous wavelet transform described above. Step 2. Perform a first-level decomposition on the vibration signal in a certain direction to obtain the low-frequency and high-frequency components. Step 3. Perform a second-level decomposition on the low-frequency and high-frequency components respectively to obtain four sub-bands. Step 4. Repeat step 3, that is, perform a third-level decomposition on the low-frequency and high-frequency components obtained in step 3 respectively to obtain eight sub-bands. Step 5. Obtain the wavelet packet coefficients or wavelet coefficients of each sub-band.

[0115] In this embodiment of the disclosure, the mother wavelet function used in the three-layer adaptive wavelet packet decomposition can be a normal, unmodified function, such as the one described above. As the mother wavelet function.

[0116] The decomposition of vibration signals in three directions using three-level adaptive wavelet packet decomposition mainly involves wavelet transform to obtain a series of wavelet coefficients. For example, taking the vibration signal x(t) in the X direction as an example, the wavelet coefficients in the X direction can be obtained using the following formula.

[0117]

[0118] in, This represents the conjugate of the basis wavelet function.

[0119] In this embodiment, the method is not limited to three-layer adaptive wavelet packet decomposition; for example, two-layer, four-layer, or five-layer adaptive wavelet packet decomposition can also be used. Too many decomposition layers can lead to information loss in the vibration signal and increase computational load. Too few decomposition layers will result in unsatisfactory noise reduction. This embodiment selects three layers, achieving a good balance between noise reduction effect and computational load. This results in a high signal-to-noise ratio and a low root mean square error (RMSE) in the vibration signal after wavelet packet denoising, while also reducing the running time.

[0120] In this embodiment, three-level adaptive wavelet packet decomposition is performed on the time-domain signals in the X, Y, and Z directions. Wavelet packet decomposition is an extension of wavelet transform; it can not only decompose the low-frequency components like wavelet transform, but also further subdivide the high-frequency components, thus providing more refined time-frequency analysis. The three-level decomposition minimizes computational complexity and redundant information while preserving the characteristics of the vibration signal. It provides sufficient resolution to distinguish between vibration signals and noise.

[0121] In this embodiment, "adaptive" in "adaptive wavelet packet decomposition" refers to the adaptive selection of the noise filtering threshold. Traditional wavelet packet denoising is based on filtering higher sub-band wavelet coefficients using a fixed threshold. However, the signal-to-noise ratio (SNR) of a vibration signal may differ at different times (i.e., it is dynamic). If a fixed threshold is used, it may lead to the misidentification of a useful signal when the noise intensity increases. Therefore, using a fixed threshold is unreasonable. This embodiment can adjust the noise filtering threshold in a timely manner. The adaptive wavelet packet denoising method provided in this embodiment is a denoising technique that dynamically updates the noise filtering threshold of wavelet coefficients. By examining the SNR of the vibration signal, a reasonable noise filtering threshold for the wavelet coefficients is constructed. Larger wavelet coefficients in the vibration signal are retained, while signals below the noise filtering threshold are removed; that is, wavelet coefficients below the noise filtering threshold are forcibly reset to zero.

[0122] For example, firstly, the signal-to-noise ratio (SNR) of the vibration signal in a certain direction is calculated in real time based on the statistical characteristics of the vibration signal in that direction; secondly, within a given interval, the noise filtering threshold of the corresponding adaptive wavelet coefficients is calculated based on the SNR; finally, by comparing the wavelet coefficients with their adaptive noise filtering thresholds, useful signals are retained, thereby dynamically adjusting the noise filtering threshold according to signal changes to improve the accuracy of noise reduction.

[0123] For example, the intensity of noise can be effectively quantified by analyzing the local variance of each wavelet coefficient. This method can capture local features in the signal, helping to identify the distributional differences between noise and signal. Furthermore, the multi-scale nature of wavelet decomposition allows local variance to be evaluated at different frequency components, aiding in the accurate differentiation between noise and signal.

[0124] In this embodiment of the disclosure, local variance refers to the variance calculated within a specific local region of the vibration signal, reflecting the degree of variation in the data within that region. In wavelet analysis, local variance can be used to measure the fluctuation of wavelet coefficients within a specific interval, thereby helping to determine the intensity of noise. Since signals may contain different characteristics and noise, this local calculation method can adapt to the non-stationarity of signals, providing more targeted analysis.

[0125] For example, a window size K (K is a positive integer greater than 1) is set, which determines the range of local variance, and then the local mean u of the wavelet coefficients is calculated. k' ,Right now:

[0126]

[0127] In the above formula, ω[n'+i'] represents the n'+i'th wavelet coefficient of the corresponding layer, where n' is a positive integer greater than or equal to 1; i' is an integer greater than or equal to 0 and less than or equal to K-1; k' represents the k'th layer of the wavelet packet adaptive decomposition used, where k' is a positive integer greater than or equal to 1. Taking three layers as an example, k' = 3.

[0128] Then calculate the local variance of the corresponding layer. For example, calculate using the following formula:

[0129]

[0130] After calculating the local variance of each wavelet coefficient at each layer, the median value is taken as the noise intensity, i.e.:

[0131]

[0132] For example, suppose the local variances of the wavelet coefficients of the vibration signal in each layer are respectively Then the median of the calculated local variance so Then, a noise filtering threshold is set based on the noise intensity. In this embodiment, the noise filtering threshold λ can be set using the following formula:

[0133]

[0134] In the above formula, d is a constant that can take any value between 1 and 2.

[0135] Similarly, different noise intensities are obtained based on the local variance of each wavelet coefficient, and different noise intensities will yield corresponding noise filtering thresholds.

[0136] By calculating the local variance, the intensity of noise in different frequency bands can be estimated. Based on the noise intensity estimation results, an appropriate noise filtering threshold is selected. The selection of the noise filtering threshold can preserve the main characteristics of the signal while removing most of the noise.

[0137] Wavelet coefficients above the noise filtering threshold are retained, while those below the threshold are forcibly set to 0. This means that wavelet coefficients above the threshold are considered the main components of the vibration signal, while those below are considered noise. Then, wavelet reconstruction (inverse wavelet transform) is performed on all wavelet coefficients. The reconstruction process is the inverse of wavelet packet decomposition; by merging wavelet coefficients layer by layer, the denoised vibration signal is finally obtained.

[0138] For example, the clean vibration signal x(t) in the X direction can be reconstructed using the following formula:

[0139]

[0140] In the above formula:

[0141]

[0142] Here, ψ(t) is the mother wavelet function used in wavelet packet decomposition, and ψ(ω) is the Fourier transform of ψ(t).

[0143] The above process can yield clean and accurate vibration signals, providing a reliable foundation for subsequent signal analysis and processing.

[0144] It should be noted that there are many methods for signal denoising, and the specific method chosen should be determined based on the characteristics of the actual signal. For non-stationary signals such as vibration signals, any one or a combination of the following denoising methods can be used: Moving average method; Savitzky-Golay method (a least squares-based convolution fitting algorithm); Median absolute deviation method (MAD); Finite Impulse Response (FIR) filter; Infinite Impulse Response (IIR) filter; Wavelet packet denoising method, etc.

[0145] The embodiments disclosed herein employ wavelet packet denoising, which achieves optimal denoising performance, with the best signal-to-noise ratio and RMSE.

[0146] In an exemplary embodiment, the method provided in this disclosure further includes: acquiring vibration sample signals of a sample object in at least two directions; performing continuous wavelet transform on the vibration sample signals of the sample object in each direction to obtain time-frequency sample images of the sample object in each direction; inputting the time-frequency sample images of the sample object in each direction into the fault identification model to obtain predicted health detection results of the sample object in each direction; obtaining the value of the model loss function of the fault identification model based on the actual health status of the sample object in each direction and the corresponding predicted health detection results; and adaptively adjusting the network structure of the fault identification model based on the value of the model loss function.

[0147] In this embodiment, vibration signals in the X, Y, and Z directions of different multi-cover surfaces at different time periods can be collected first. The vibration signals in the X, Y, and Z directions of the same multi-cover surface at the same time period are grouped together. Then, the vibration signals in the three directions of each group are subjected to noise reduction preprocessing. The vibration signals after noise reduction preprocessing are subjected to continuous wavelet transform (CWT) to obtain a set of time-frequency images of wavelets (dataset). The dataset is divided into a training set (each group of time-frequency images is used as a training sample; for distinction, the object to be detected is called the sample object, the vibration signal is called the vibration sample signal, and the time-frequency image is called the time-frequency sample image), a validation set (each group of time-frequency images is used as a validation sample; for distinction, the object to be detected is called the validation object, the vibration signal is called the vibration validation signal, and the time-frequency image is called the time-frequency validation image), and a test set (each group of time-frequency images is used as a test sample, or the time-frequency images of at least two directions of the multi-cover surface in the above embodiment).

[0148] For example, vibration signals in the X, Y, and Z directions of multiple multi-cover surfaces are collected at different times (e.g., 1000 vibration signals in the X, Y, and Z directions are collected). Then, wavelet transform is performed on each of these 1000 sets of pre-processed, denoised vibration signals in the X, Y, and Z directions. After the transform, each set will contain time-frequency images of the wavelet coefficients of the vibration signals in the X, Y, and Z directions; that is, each set contains time-frequency images for each of the X, Y, and Z directions. Of these 1000 sets (each set containing three time-frequency images), 700 sets are divided into a training set, 200 sets into a validation set, and 100 sets into a test set. The labels (i.e., the true health status) of each training sample in the training set and each validation sample in the validation set are known. For example, the true health status corresponding to the three time-frequency images in a certain set in the training set is known to be either: the multi-cover surface is faulty, or the multi-cover surface is not faulty.

[0149] During the training phase of a fault identification model, such as a CNN model, the time-frequency maps of wavelets in the X, Y, and Z directions of each group in the training set are input into the CNN model. The CNN model predicts the probability of whether a multi-cover surface has a fault based on the time-frequency maps of wavelets in the X, Y, and Z directions. For example, taking one group as an example, the CNN model will output six probability values: p(X) (representing the probability of the multi-cover surface having a fault based on the time-frequency map of wavelets in the X direction), 1-p(X) (representing the probability of the multi-cover surface not having a fault based on the time-frequency map of wavelets in the X direction), p(Y) (representing the probability of the multi-cover surface having a fault based on the time-frequency map of wavelets in the Y direction), 1-p(Y) (representing the probability of the multi-cover surface not having a fault based on the time-frequency map of wavelets in the Y direction), p(Z), and 1-p(Z). Then, the magnitudes of p(X) and 1-p(X), p(Y) and 1-p(Y), and p(Z) and 1-p(Z) are compared to obtain the predicted health detection results in the X / Y / Z directions, namely "Normal (normal, no fault in the multi-cover surface)" and "Abnormal (abnormal, fault in the multi-cover surface)". Based on the actual health status of the sample object in each direction and the corresponding predicted health detection results, the model loss function of the fault identification model is constructed, and the fault identification model is trained. After training, the trained fault identification model can be validated on a validation set until preset conditions are met, such as the model prediction accuracy reaching a preset condition, reaching the maximum number of iterations, or the model loss function value reaching a preset threshold.

[0150] For example, the model loss function or error loss function J(θ,c) used in the embodiments of this disclosure can be expressed by the following formula:

[0151]

[0152] In the above formula, g θ,c (x (i1) ) and y (i1) These are the i-th training samples x (i1) The model calculates the predicted health check results and the actual health status, where i1 is a positive integer greater than or equal to 1 and less than or equal to M, and M is a positive integer greater than or equal to M, representing the number of training samples. θ and c are parameters of the fault identification model. The purpose of training the fault identification model is to obtain the minimum value of J(θ,c). Gradient descent can be used to update the parameters θ and c.

[0153] In an exemplary embodiment, the network structure of the fault identification model is adaptively adjusted based on the value of the model loss function, including: if the decrease in the value of the model loss function within a first preset time period is lower than a preset value (the value of the preset value can be set according to actual needs, and this disclosure does not limit it, as long as it can indicate that the decrease in the model loss function is not significant) and is unstable, then the number of nodes in the fault identification model is increased; if the value of the model loss function does not reach a preset threshold within a second preset time period (the value of the preset threshold can be set according to actual scenarios, and this disclosure does not limit it), then the number of network layers in the fault identification model is increased.

[0154] Figure 3 This is a schematic diagram illustrating the change in accuracy of a fault identification model as it undergoes training iterations, according to an exemplary embodiment of this disclosure. Figure 3 The x-axis represents accuracy, and the y-axis represents the number of epochs (the number of times a complete dataset passes through the fault identification model once). The entire training and validation process of the fault identification model is as follows: Figure 3 As shown, the training accuracy gradually increases with the number of training rounds, and then stabilizes after 100 training rounds.

[0155] Figure 4 This is a schematic diagram illustrating the change of the loss function of a fault identification model as it undergoes training iterations, according to an exemplary embodiment of this disclosure. Figure 4 In the example, the horizontal axis represents the value of the model's loss function (also known as the loss value), and the vertical axis represents the number of rounds / rounds. From Figure 4 This can be viewed as the loss function (i.e., the training loss) gradually stabilizing after 100 training epochs. After 200 training epochs, the fault identification model has reached its optimal performance. If the loss function fails to converge or continues to fluctuate, this may be considered as gradient descent overshoot or incorrect initial hyperparameter settings.

[0156] Figure 5 Taking a CNN model as an example, a network structure diagram of an exemplary fault identification model is given. Figure 5As shown, the CNN model includes an input layer 51, a convolutional layer 52, batch normalization 53, a max pooling layer 54, first residual blocks 55, 56, and 57, second residual blocks 58, 59, and 510, a convolutional layer 511, an activation function 512, residual connections 513, convolutional layers 514, residual connections 515, global average pooling 516, and a fully connected layer and activation function 517. The residual blocks are added to the convolutional layers through residual skip connections to deepen the model while ensuring the preservation of previously recognizable features. It consists of multiple layers, such as 3×3 convolutional layers with batch normalization and ReLU. The internal structure of the CNN model can be added, removed, or rearranged as needed, and is not limited to the examples described above. The CNN model is trained using the backpropagation algorithm, minimizing the loss function by continuously adjusting the weights and biases in the convolutional layers.

[0157] In this embodiment of the disclosure, a fault identification model with adaptive structural adjustment, such as a CNN model, is established. Specifically, the network structure of the fault identification model, such as the CNN model, can be adaptively adjusted according to the changing trend of the model's loss function.

[0158] In some embodiments, if the loss function of a fault identification model, such as a CNN model, does not decrease significantly and is unstable, it may indicate that the current fault identification model lacks sufficient complexity and cannot effectively capture the features of the data. In this case, increasing the number of nodes in the network (i.e., increasing the number of neurons in each layer) can better fit the data. For example, as described above... Figure 5 For example, a basic fully connected layer might contain 128 nodes, which can be increased to 256 nodes or even more to enhance the expressive power of the CNN model. If increasing the number of nodes in the fully connected layer results in a faster decrease in the loss function, a more stable loss function, or an improvement in accuracy, then this measure is effective.

[0159] In other embodiments, if the loss function of the fault detection model, such as a CNN model, cannot approach a preset threshold, the number of network layers is increased. For example, as described above... Figure 5 For example, new layers (such as new convolutional layers) can be added between convolutional layers, which can help fault detection models such as CNN models extract more intermediate features from the data. Especially for complex image data, adding convolutional layers can better capture features at different scales.

[0160] This embodiment employs a fault identification model, such as a CNN model, whose structure can be adaptively adjusted, to extract features and identify faults from the CWT time-frequency image of vibration signals. Compared to other network models, this method can adaptively determine the network structure based on the characteristics of the input data, effectively avoiding the blindness of network structure selection and reducing network training time.

[0161] In an exemplary embodiment, the method provided in this disclosure also develops a visual user interface platform. This platform can be, for example, a user-oriented interface platform developed using the Python language, which has functions such as data uploading, result analysis and display, visualization, and data saving. Through the development of this user interface platform, the results of a multi-faceted health check (i.e., health test results) of the object to be detected can be presented intuitively in the form of data and images.

[0162] For example, such as Figure 6 The user interface platform described above can be a visual interface of an MPC Health Detection System 600. It can be divided into two areas, 610 and 620. Area 610 contains a first control 611, a second control 612, a third control 613, and a fourth control 614. Triggering the first control 611 displays the CWT time-frequency image input to the CNN model. Triggering the second control 612 allows data uploading, specifically uploading vibration signals collected by vibration sensors to the MPC health detection system or fault detection system. Triggering the third control 613 allows analysis of the uploaded vibration signals to obtain time-frequency images of each multi-cover surface collected in three directions at different time periods. Based on these time-frequency images, the health detection results of each multi-cover surface are obtained using the aforementioned method, and the health detection results of each multi-cover surface at different time periods are displayed. For example… Figure 6 The data displays the health check results of MPC2 at different time periods. Checkmarks indicate normal results for the corresponding time period, while crosses indicate abnormal results. At the bottom of area 610, it shows that 45 of the MPC2 health check results were normal, 48 were abnormal, and the total was 93. Triggering the fourth control 614 will save the analyzed data.

[0163] Region 620 displays the conclusion based on the analysis results in Region 610. Here, it is assumed that the conclusion is uncertain, with a risk index of approximately 6 / 10, meaning the MPC2 health check result is the ratio of the number of abnormalities to the total. It shows a normal percentage of 48.4% and an abnormal percentage of 51.6%, presented in pie chart 621. Because the normal and abnormal percentages are quite close, the conclusion is uncertain, meaning it cannot be determined whether the multi-cover surface is faulty.

[0164] When multiple multi-cover surfaces are inspected using the above method, they can be sorted by the percentage of abnormality to determine the priority of handling each multi-cover surface that may have a fault. That is, the multi-cover surface with the higher percentage of abnormality is given priority for troubleshooting.

[0165] Area 620 also includes a fifth control 622. Triggering the fifth control 622 can exit the display interface of the health test results of MPC2.

[0166] This disclosed embodiment combines deep learning and information fusion technology to fill the gap in the field of multi-coverage intelligent health inspection, effectively improves the efficiency and accuracy of health status inspection, reduces the cost of multi-coverage fault detection and maintenance, helps management departments to grasp the health status of multi-coverage in a timely manner, and provides guidance for equipment maintenance.

[0167] The fault detection method provided in this disclosure can realize non-invasive short-distance fault detection of multiple covers on urban roads. The equipment is portable and can collect data in a semi-automatic / automatic manner with simple deployment and settings on the sidewalk. It does not affect driving safety and does not require road closures or uncovering of multiple covers. It can identify problematic multiple covers through vibration signals in on-site environments such as traffic noise, thereby improving work efficiency and reducing the impact on traffic.

[0168] The fault detection method provided in this disclosure is a non-invasive, multi-coverage, short-range intelligent health inspection method for cities based on an improved DS evidence theory. It addresses the problems of low efficiency, low accuracy, high cost, and interference with road traffic inherent in related diagnostic methods. Innovations are made in several aspects, including feature extraction, pattern recognition, and decision fusion. To reduce the impact of boundary effects on the diagnostic performance of fault identification models such as CNN models, the mother wavelet function in the traditional continuous wavelet transform is improved, thereby enhancing the recognition accuracy of fault identification models such as CNN models. Furthermore, the traditional DS evidence theory is improved by assigning different weight coefficients according to the importance of different evidence bodies, effectively avoiding the weakening of feature conflicts between different evidence focuses.

[0169] Figure 7 The embodiment uses multiple surfaces as the object to be detected and collects vibration signals in the X / Y / Z directions of the multiple surfaces. The fault identification model used is a CNN model for illustration. For example, this method can be a non-invasive urban multi-surface short-distance intelligent health inspection method based on improved DS evidence theory. Figure 7 The methods provided in the embodiments can be executed by any electronic device, such as a terminal and / or a server. Figure 7 As shown, the method provided in this disclosure embodiment may include the following steps.

[0170] Step S71: Obtain vibration signals of the multiple covers in the X / Y / Z directions.

[0171] For example, a vibration sensor can collect vibration signals from multiple surfaces in the X, Y, and Z directions at a short distance using a non-direct contact method, thus obtaining the system's time-domain dataset. The vibration sensor then sends these signals to an oscilloscope for display. Users can select the appropriate vibration signal for further analysis based on the waveforms displayed on the oscilloscope. Alternatively, when the vibration sensor receives a trigger signal from a photoelectric sensor, it can also send the collected vibration signals from multiple surfaces in the X, Y, and Z directions to an oscilloscope for display. Users can then select the appropriate vibration signal for further analysis based on the waveforms displayed on the oscilloscope.

[0172] Step S72: Noise reduction processing is performed on the vibration signals in the X, Y, and Z directions respectively.

[0173] For example, the specific method for denoising data in S72 can be as follows: First, the time-domain signal in the X / Y / Z directions is decomposed using a three-layer adaptive wavelet packet decomposition method; second, the local variance of each wavelet coefficient is calculated to estimate the noise intensity, and then an appropriate noise filtering threshold is selected; finally, wavelet coefficients with values ​​greater than the noise filtering threshold are retained for reconstruction to obtain a clean vibration signal.

[0174] Step S73: Perform continuous wavelet transform on the noise-reduced vibration signals in the X, Y, and Z directions respectively to obtain time-frequency images in the X, Y, and Z directions.

[0175] For example, after performing continuous wavelet transform, a CWT time-frequency image set is obtained (the CWT time-frequency image set includes multiple time-frequency images).

[0176] For example, the specific method for performing continuous wavelet transform on the denoised data in S73 is as follows: based on the traditional mother wavelet function, waveform adjustment is added to adjust the decay rate of the basis wavelet function, thus solving the boundary effect present in the traditional wavelet transform. Waveform adjustment refers to changing the shape or parameters of the mother wavelet function to better adapt it to the local characteristics of the signal, thereby reducing the influence of boundary effects.

[0177] Step S74: Divide the time-frequency image into training samples, validation samples, and test samples.

[0178] Step S75: Build a CNN model by training and validating the CNN model using training samples and validation samples respectively.

[0179] For example, training and validation can be performed on time-frequency images in the X / Y / Z directions of training samples and validation samples, respectively.

[0180] In this embodiment of the disclosure, the method for establishing a CNN model with adaptive structural adjustment in step S75 can be as follows: adaptively adjusting the network structure according to the changing trend of the model's loss function. If the loss function does not decrease significantly and is unstable, the number of network nodes is increased; if the loss function cannot approach a preset threshold, the number of network layers is increased.

[0181] Step S76: Input the test sample into the CNN model to obtain the detection results of the multiple surfaces in the X / Y / Z directions (i.e., the initial health detection results).

[0182] Step S77: Utilize the improved DS evidence theory to perform decision fusion on the detection results (or diagnostic results) of multiple surfaces in the X / Y / Z directions to obtain the final comprehensive diagnostic conclusion or the final comprehensive detection result (i.e., health detection result).

[0183] Step S78: Display the final comprehensive test results on the user interface platform.

[0184] Specifically, a user interface platform will be developed to display the above analysis and test results (i.e., health test results) on the terminal in the form of data or images.

[0185] Figure 7 Other aspects of the embodiments can be found in the other embodiments described above, and will not be repeated here.

[0186] The fault detection method provided in this disclosure is a non-invasive short-distance remote fault diagnosis method for multi-part covers in urban areas based on improved Dempster-Shafer (DS) evidence theory. This method utilizes a portable vibration sensor to collect vibration signals in the X, Y, and Z directions from a pedestrian's side at a short distance, obtaining raw time-domain data. The vibration signals in the X, Y, and Z directions are preprocessed with noise reduction. The preprocessed signals are then subjected to continuous wavelet transform (CWT) to obtain a wavelet time-frequency image set (containing multiple time-frequency images). The dataset is divided into a training set, a validation set, and a test set according to a certain ratio (e.g., 7:2:1). A two-dimensional convolutional neural network (not limited to a CNN model; any machine learning model capable of classification and recognition (e.g., whether a fault exists)) model is established, and the network model is trained and tested using the wavelet time-frequency image set in the X, Y, and Z directions to obtain diagnostic results in the three directions. A decision fusion algorithm based on an improved DS evidence theory is also proposed to fuse the diagnostic results from the three directions mentioned above to obtain the final comprehensive detection result. In addition, a visualization user platform has been developed to display the comprehensive examination results, i.e., the final comprehensive detection result.

[0187] The fault detection method provided in this disclosure combines information fusion and deep learning technologies to perform early diagnosis of the health status of multiple surfaces covered by road, improving inspection efficiency and accuracy, saving labor costs, and filling a gap in the field of early health inspection of multiple surfaces covered by road. This method is simple and easy to operate, and can be used on pedestrian walkways without affecting road traffic safety.

[0188] The fault detection method provided in this disclosure simultaneously collects vibration signals or vibration data in three directions (X, Y, and Z) and uses them as fault data for feature extraction. Compared with image vision methods, it can better reflect the changing patterns of excitation signals inside multiple covers, which helps to detect internal faults in the system.

[0189] The fault detection method provided in this disclosure uses a CNN model with an adaptively adjustable structure to extract features and identify faults from the CWT time-frequency image of vibration signals. Compared with other network models, this method can adaptively determine the network structure according to the characteristics of the input data, effectively avoiding the blindness of network structure selection and reducing the network training time.

[0190] The fault detection method provided in this disclosure adopts an information fusion method based on the improved DS evidence theory, which fuses the diagnostic results in three directions at the decision level. Compared with the algorithm before the improvement, it can effectively improve the accuracy of diagnosis.

[0191] The fault detection method provided in this disclosure can be applied to the health status detection of multiple covers in any region. It can also be applied to the fault detection of other equipment, such as the health status detection of other types of drain covers, bridges, roads, etc. The solutions provided in this disclosure relate to AI, big data, computing / information technology, and data science.

[0192] Figure 8 This is a schematic diagram of a fault detection device according to an exemplary embodiment of this disclosure. In the exemplary embodiment, the fault detection device may be located within an electronic device or implemented via an electronic device. Figure 8 As shown, the fault detection device 800 provided in this embodiment may include a receiving unit 810 and a processing unit 820.

[0193] The receiving unit 810 is used to acquire vibration signals of the object to be detected in at least two directions.

[0194] The processing unit 820 is used to obtain the initial health detection results of the object under test in at least two directions based on the vibration signals of the object under test in at least two directions.

[0195] The processing unit 820 is also used to fuse the initial health detection results of the object to be detected in at least two directions to determine the health detection result of the object to be detected.

[0196] Figure 8 Other aspects of the embodiments can be found in the above embodiments, and will not be repeated here.

[0197] Figure 9 This is a schematic diagram of the structure of a fault detection system according to an exemplary embodiment of this disclosure. Figure 9 As shown, the fault detection system 900 provided in this embodiment includes a vibration sensor 920 and a fault detection device 800.

[0198] The object to be detected 910 is within the coverage area of ​​the vibration sensor 920, and there is a distance r between the object to be detected 910 and the vibration sensor 920, where r is a real number greater than 0. The vibration sensor 920 is used to detect the vibration signals of the object to be detected 910 in at least two directions.

[0199] The fault detection device 800 includes a receiving unit 810. The receiving unit 810 is used to acquire vibration signals of the object 910 under test detected by the vibration sensor 920 in at least two directions.

[0200] Other aspects of the fault detection device 800 can be found in the other embodiments described above, and will not be repeated here.

[0201] In an exemplary embodiment, the object to be detected 910 includes multiple covers. The fault detection system 900 may further include a photoelectric sensor for sending a trigger signal to instruct the acquisition of vibration signals from the multiple covers in at least two directions collected by the vibration sensor 920 when a vehicle is detected passing over the multiple covers.

[0202] In an exemplary embodiment, the fault detection system 900 may further include: a support plate fixed to the ground, and a vibration sensor 920 fixed to the support plate.

[0203] Figure 10 The illustration shows an application scenario of a fault detection system according to an embodiment of the present disclosure. Figure 10 The example uses an object to be tested as a multi-faceted cover and an aluminum plate as a support plate for illustration.

[0204] like Figure 10 As shown, a multi-faceted surface 102 is provided on the surface of road 101. An aluminum plate 108 is first fixed at an appropriate position on the sidewalk next to road 101, and then a vibration sensor 103, such as an accelerometer, is fixed to the aluminum plate 108. This ensures that the vibration sensor 103 is positioned appropriately on the sidewalk, i.e., at a suitable distance from the multi-faceted surface 102, allowing it to collect vibration signals from the multi-faceted surface 102 in the X, Y, and Z directions. For example, the X-axis of the vibration sensor 103 points towards the multi-faceted surface 102, and the Y-axis is parallel to the edge of the sidewalk. In actual operation, the X-axis points (positive / negative) towards the multi-faceted surface, the Y-axis is perpendicular to the X-axis on the same road surface, and the Z-axis measures the vertical vibration signal.

[0205] The oscilloscope 107 can be connected to both a power station 105 and an amplifier 106. The power station 105 supplies power to both the oscilloscope 107 and the amplifier 106. The amplifier 106 is connected to the vibration sensor 103. The amplifier 106 can receive vibration signals from the multi-faceted surface 102 in the X, Y, and Z directions sent by the vibration sensor 103, amplify the vibration signals in the X, Y, and Z directions of the multi-faceted surface 102 respectively, and transmit the amplified vibration signals in the X, Y, and Z directions of the multi-faceted surface 102 to the oscilloscope 107 for display.

[0206] Optionally, a photoelectric sensor 104 can be deployed at an appropriate location on the sidewalk. This photoelectric sensor 104 is used to detect whether a vehicle, such as a car, has passed over the multi-cover 102. When a vehicle is detected passing over the multi-cover 102, the photoelectric sensor 104 can send a trigger signal to a vibration sensor 103, which then sends the detected vibration signal to an amplifier 106 for display on an oscilloscope 107. For example, the photoelectric sensor 104 is positioned close to the edge of the road 101, and there are no other objects obstructing the detection target, such as the vehicle passing over the multi-cover 102, to avoid false detections caused by pedestrians crossing the sidewalk. The user can observe the waveform of the vibration signal displayed on the oscilloscope 107 and select the appropriate vibration signal for subsequent fault detection analysis.

[0207] It should be noted that this disclosure is not limited to using photoelectric sensor 104 to detect whether a vehicle has driven over the multi-cover surface 102. For example, pressure sensor, infrared sensor, etc. can also be used for detection.

[0208] In some embodiments, the vibration sensor 103 may have a storage device in which the collected vibration signals can be stored. When fault analysis is required, the corresponding vibration signal can be retrieved from the storage device for fault analysis.

[0209] Furthermore, embodiments of this disclosure also provide an electronic device, including: one or more processors; and a memory configured to store one or more programs, which, when executed by the one or more processors, cause the electronic device to perform the method as described in any embodiment of this disclosure.

[0210] The electronic device includes a processor that can call and run computer programs from memory to implement the methods described in the embodiments of this disclosure.

[0211] Optionally, the electronic device may further include a memory. The processor can retrieve and run computer programs from the memory to implement the methods in any embodiment of this disclosure.

[0212] The memory can be a separate device independent of the processor, or it can be integrated into the processor.

[0213] Optionally, the electronic device may also include a transceiver, which the processor can control to communicate with other devices. Specifically, it can send information or data to other devices or receive information or data sent by other devices.

[0214] Optionally, the processor, memory, and transceiver can communicate bidirectionally with each other via a communication bus.

[0215] It should be understood that the processor in this embodiment of the disclosure may be an integrated circuit chip with signal processing capabilities. In implementation, the steps of the above method embodiments can be completed by integrated logic circuits in the processor's hardware or by software instructions.

[0216] The aforementioned processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this disclosure. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this disclosure can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above methods.

[0217] It is understood that the memory in the embodiments of this disclosure can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced Synchronous DRAM (ESDRAM), Synchlink DRAM (SLDRAM), and Direct Rambus RAM (DR RAM). It should be noted that the memory used in the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory. It should be understood that the above-described memory is exemplary and not limiting.

[0218] Furthermore, embodiments of this disclosure also provide a computer-readable storage medium storing a computer program that, when run on a computer, causes the computer to perform the methods described in any embodiment of this disclosure.

[0219] This disclosure also provides a computer-readable storage medium for storing a computer program. This computer program causes a computer to perform corresponding processes in the various methods of the embodiments of this disclosure; for brevity, these will not be elaborated upon here.

[0220] Furthermore, embodiments of this disclosure also provide a computer program product, including a computer program that, when executed by a processor, implements the methods described in any embodiment of this disclosure.

[0221] This disclosure also provides a computer program product, including computer program instructions. These computer program instructions cause a computer to execute corresponding flows in the various methods of the embodiments of this disclosure; for simplicity, they will not be described in detail here. This disclosure also provides a computer program. When this computer program is run on a computer, it causes the computer to execute corresponding flows in the various methods of the embodiments of this disclosure; for simplicity, they will not be described in detail here.

Claims

1. A fault detection method, characterized in that, include: Acquire vibration signals of the object to be detected in at least two directions; Based on the vibration signals of the object under test in at least two directions, the initial health detection results of the object under test in at least two directions are obtained; The initial health detection results of the object to be detected in at least two directions are fused to determine the health detection result of the object to be detected.

2. The method as described in claim 1, characterized in that, The initial health detection results of the object to be detected in at least two directions are fused to determine the health detection result of the object to be detected, including: The initial health test results of the object to be tested in each direction are used as the corresponding evidence. Calculate the similarity between pieces of evidence; Based on the similarity between evidence pieces, the mutual support between evidence pieces is obtained; The weight coefficient of each piece of evidence is obtained based on the degree of mutual support between them. Based on the evidence and its corresponding weighting coefficients, the health test results of the object to be tested are determined.

3. The method as described in claim 1, characterized in that, Based on the vibration signals of the object under test in at least two directions, initial health detection results of the object under test in at least two directions are obtained, including: Continuous wavelet transform is performed on the vibration signals of the object to be detected in each direction to obtain time-frequency images of the object to be detected in each direction. The time-frequency images of the object to be detected in each direction are input into the fault identification model to obtain the initial health detection results of the object to be detected in each direction.

4. The method as described in claim 3, characterized in that, Perform continuous wavelet transform on the vibration signals of the object under test in each direction to obtain time-frequency images of the object under test in each direction, including: Determine the mother wavelet function, including waveform adjustment parameters; The basis wavelet function is determined based on the mother wavelet function, which includes waveform adjustment parameters, wherein the waveform adjustment parameters are used to adjust the decay rate of the basis wavelet function. Based on the basis wavelet function, continuous wavelet transform is performed on the vibration signal of the object to be detected in each direction to obtain the wavelet coefficients of the object to be detected in each direction. Based on the wavelet coefficients of the object to be detected in each direction, the time-frequency images of the object to be detected in each direction are determined.

5. The method as described in claim 4, characterized in that, Mother wavelet function φ including waveform adjustment parameters α (t) is expressed by the following formula: Where μ is the normalization coefficient; α is the waveform adjustment parameter; ξ is the damping ratio parameter, which is a real number greater than 0 and less than 1; t represents time; and ω represents the angular frequency.

6. The method as described in claim 3, characterized in that, Perform continuous wavelet transform on the vibration signals of the object under test in each direction to obtain time-frequency images of the object under test in each direction, including: The vibration signals of the object to be detected in each direction are subjected to noise filtering to obtain the vibration signals of the object to be detected in each direction after noise removal. After removing noise, the vibration signals of the object under test in each direction are subjected to continuous wavelet transform to obtain time-frequency images of the object under test in each direction.

7. The method as described in claim 6, characterized in that, Noise filtering is performed on the vibration signals of the object under test in various directions to obtain the noise-removed vibration signals of the object under test in various directions, including: The vibration signal of the object to be detected in each direction is decomposed using a three-layer adaptive wavelet packet decomposition method to obtain wavelet coefficients. Calculate the local variance of each wavelet coefficient; The noise intensity of each wavelet coefficient is determined based on the local variance of each wavelet coefficient; The noise filtering threshold for each wavelet coefficient is determined based on the noise intensity of that wavelet coefficient. Wavelet coefficients with values ​​greater than the corresponding noise filtering threshold are retained for reconstruction to obtain the vibration signals of the object under test in various directions after noise removal.

8. The method as described in claim 3, characterized in that, Also includes: Acquire vibration sample signals of the sample object in at least two directions; Continuous wavelet transform is performed on the vibration sample signals of the sample object in each direction to obtain time-frequency sample images of the sample object in each direction. The time-frequency sample images of the sample object in each direction are input into the fault identification model to obtain the predicted health detection results of the sample object in each direction. Based on the actual health status of the sample object in each direction and the corresponding predicted health detection results, the value of the model loss function of the fault identification model is obtained; Based on the value of the model loss function, the network structure of the fault identification model is adaptively adjusted.

9. The method as described in claim 8, characterized in that, Based on the value of the model loss function, the network structure of the fault identification model is adaptively adjusted, including: If the value of the model loss function decreases by less than a preset value and is unstable within a first preset time period, then the number of nodes in the fault identification model is increased. If the value of the model loss function does not reach the preset threshold within the second preset time period, the number of network layers of the fault identification model is increased.

10. The method as described in claim 1, characterized in that, The object to be detected includes multiple covered surfaces; The process of acquiring vibration signals of the object to be detected in at least two directions includes: Vibration signals in at least two directions are acquired by vibration sensors as a vehicle passes over the multi-cover surface.

11. The method as described in claim 10, characterized in that, Acquiring vibration signals in at least two directions, collected by vibration sensors, as a vehicle passes over the multi-cover surface, including: Upon receiving a trigger signal from the photoelectric sensor, the vibration signals in at least two directions collected by the vibration sensor are acquired. The trigger signal of the photoelectric sensor is used to indicate that a vehicle has passed over the multi-cover surface.

12. The method as described in claim 10 or 11, characterized in that, The multiple covers are installed on the road, and the vibration sensors are deployed on the sidewalk, with the multiple covers within the coverage area of ​​the vibration sensors.

13. A fault detection device, characterized in that, include: A receiving unit is used to acquire vibration signals of the object to be detected in at least two directions; The processing unit is used to obtain the initial health detection results of the object under test in at least two directions based on the vibration signals of the object under test in at least two directions; The processing unit is further configured to fuse the initial health detection results of the object to be detected in at least two directions to determine the health detection result of the object to be detected.

14. A fault detection system, characterized in that, include: A vibration sensor is used to detect vibration signals of the object to be detected in at least two directions, provided that the object to be detected is within the coverage area of ​​the vibration sensor and there is a distance between the object to be detected and the vibration sensor. The fault detection device as described in claim 13, wherein the receiving unit is used to acquire the vibration signals of the object to be detected detected by the vibration sensor in at least two directions.

15. The system as described in claim 14, characterized in that, The object to be detected includes multiple covered surfaces; wherein, the system further includes: A photoelectric sensor is used to send a trigger signal to indicate the acquisition of vibration signals of the multi-cover surface in at least two directions collected by the vibration sensor when a vehicle is detected passing over the multi-cover surface.

16. The system as described in claim 14, characterized in that, Also includes: A support plate is fixed to the ground, and the vibration sensor is fixed to the support plate.

17. An electronic device, characterized in that, include: One or more processors; A memory configured to store one or more programs that, when executed by one or more processors, cause the electronic device to perform the method of any one of claims 1 to 12.

18. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is run on a computer, it causes the computer to perform the method of any one of claims 1 to 12.

19. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 12.

Images