Multi-modal spatio-temporal alignment and three-layer fusion reasoning based power inspection robot fault diagnosis method

Abstract

The application discloses a power inspection robot fault diagnosis method based on multi-modal space-time alignment and three-layer fusion reasoning, and particularly relates to the technical field of power equipment inspection fault diagnosis, and comprises the following steps: acquiring an inspection package read by an edge computing unit of an inspection robot from an intelligent sensing system, generating a sampling word according to a sampling order, wherein the sampling word is composed of an Euler angle quantization code, an electrical component index and a cable section index, and writing a sign inversion position of an electrical difference value of adjacent sampling words as a doubtful word; restoring robot Euler angles based on the doubtful word, generating a rotation matrix of a vehicle body coordinate to a station coordinate through an Euler angle coordinate transformation algorithm, and converting a sampling direction of the doubtful word into a station observation vector; writing multi-modal abnormalities into the doubtful word with postures, components and cable attributions, then inversely deducing a source action at an edge end, and re-sampling component electrical responses and cable echoes, and finally aligning, direction checking and electrical connection checking responses before and after the action.

Description

Technical Field

[0001] This invention relates to the field of power equipment inspection and fault diagnosis technology, and more specifically, to a power inspection robot fault diagnosis method based on multimodal spatiotemporal alignment and three-layer fusion reasoning. Background Technology

[0002] When existing power inspection robots perform fault diagnosis, they generally obtain on-site data through sensing methods such as infrared imaging, visible light imaging, partial discharge detection, and gas sampling. They then combine time synchronization, robot positioning, equipment ledgers, and fusion models to determine whether electrical equipment is overheating, discharging, or has insulation abnormalities. In the inspection of rows of switchgear, cable terminals and bus joints in indoor substations, robots need to move along narrow passages according to the inspection rhythm. The time available for stopping and verifying at the edge is limited, and it is difficult to upload all the original data in time for manual correction. Under these operating conditions, relying solely on the acquisition time and robot position to merge multimodal anomalies can easily lead to anomalies from different sources being grouped into the same fault object. On-site manifestations include: after the robot moves to the side, the infrared hot spot changes with the position of the reflected light; the partial discharge peak changes with distance and points towards the adjacent cabinet; the gas response falls into the next sampling segment due to the lag in intake; and the suspected defects in visible light disappear after the supplementary light angle is changed. However, the diagnostic results still merge these anomalies into the same electrical component fault. The reason for this problem is that time synchronization and spatial positioning can only prove that multiple types of data were obtained at similar times and locations, but cannot prove that infrared, visible light, partial discharge and gas anomalies all come from the same electrical fault object. The technical problem this application aims to solve is: how to avoid the incorrect binding of multimodal anomaly evidence to the same electrical fault object under the conditions of mobile data collection by inspection robots and real-time diagnosis at the edge. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a fault diagnosis method for power inspection robots based on multimodal spatiotemporal alignment and three-layer fusion inference. This method first writes the multimodal anomalies into suspicious words containing attitude, component, and cable attribution, then reverse-engineers the source verification action from the edge end and re-samples the component electrical response and cable echo, and finally aligns, verifies the direction, and verifies the electrical connection of the responses before and after the action, thereby solving the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a fault diagnosis method for power inspection robots based on multimodal spatiotemporal alignment and three-layer fusion inference, comprising: S1. Obtain the inspection package read by the edge computing unit of the inspection robot from the intelligent sensing system, generate sampling words according to the sampling order, the sampling words are composed of Euler angle quantization code followed by electrical component index and then cable segment index, and reverse the position of the electrical measurement difference sign of adjacent sampling words to write the suspicious point word. S2. Based on the suspicious words, restore the robot's Euler angles, generate the rotation matrix from the vehicle coordinates to the station coordinates through the Euler angle coordinate transformation algorithm, convert the sampling direction of the suspicious words into the station observation vector, and then subtract the direction corresponding to the electrical component index to generate the source verification action word. S3. Drive the inspection robot to change the sampling posture according to the source verification action word. At the action completion position, the electrical performance testing device obtains the component's electrical response, the electrical fault detection device obtains the cable echo, and writes it as the re-sampling word according to the source verification action number. S4. Perform the neural dynamic time warping algorithm on the source verification sequence pair composed of the suspicious word and the re-collected word. The neural encoder generates response codes, writes the absolute value of the difference between response codes at the same index position into the grid cost table, and writes the first item into the path table in ascending order of the cumulative cost of the three predecessor paths. Backtrack to generate the source verification alignment chain. S5. Calculate the response difference direction before and after the verification action along the verification source alignment chain. Multiply the response difference direction with the source direction of the verification action word bit by bit. Delete suspicious words with negative products from the verification source alignment chain. The retained suspicious words are checked for consistency between electrical detection rules, electrical connection continuity and fault semantics. Output the inspection robot fault diagnosis result.

[0005] In a preferred embodiment, S1 includes: S11. After obtaining the inspection package, the edge computing unit uses the sampling order as the write key to generate the Euler angle quantization code by dividing the robot's Euler angle by the angle resolution, and writes the Euler angle quantization code, part number and line segment number into the sampling word in a fixed length and bit width. The inspection package refers to the field recording frame that the edge computing unit reads from the intelligent sensing system and encapsulates within a sampling sequence. The recording frame is indexed by the sampling sequence and contains robot posture, electrical object, line section, electrical performance test reading and fault detection reading, which are used to generate sampling words. The Euler angles of a robot refer to the vehicle body attitude angles output by the attitude sensing unit of the inspection robot. The vehicle body attitude angles are used to describe the rotational relationship between the robot's vehicle body coordinates and its coordinates within the station, and serve as the source for calculating the Euler angle quantization code. The part number is a unique index assigned to the electrical component being inspected in the electrical equipment ledger. The part number is used to read the electrical performance test value of the same electrical component in adjacent sampling order and to keep the object attribution of the electrical performance test value consistent. The segment number is a unique index corresponding to a cable, transmission line, or network segment in the electrical connection topology. The segment number is used to read the fault detection value of the same line segment under the adjacent sampling sequence and to keep the line attribution of the fault detection value consistent. S12. Read the electrical performance test value based on the component number of the sampling word and generate the component difference symbol; read the fault detection value based on the line segment number of the sampling word and generate the line segment difference symbol; multiply the component difference symbol and the line segment difference symbol to generate the electrical test symbol bit. The electrical performance test value refers to the digital reading of electrical parameters obtained by the electrical performance testing device at the electrical component corresponding to the component number. The digital reading of electrical parameters is written into the inspection package according to the sampling order and is used to calculate the component difference sign between adjacent sampling orders. The fault detection value refers to the fault response digital reading obtained by the electrical fault detection device at the line segment corresponding to the line segment number. The fault response digital reading is written into the inspection package according to the sampling order and is used to calculate the line segment difference sign between adjacent sampling orders. S13. Under the same sampling order, read the electrical measurement symbol bit. When the product of the current electrical measurement symbol bit and the previous electrical measurement symbol bit is negative, the edge computing unit writes the sampling word corresponding to the current sampling order as the doubt word.

[0006] In a preferred embodiment, S2 includes: S21. Based on the suspicious words, read the Euler angle quantization code and sampling direction code, extract three segments of angle code according to the writing order of the Euler angle quantization code, multiply each segment of angle code by the angle resolution to obtain three segments of angle value, and restore the sampling direction code to the vehicle body sampling vector. The sampling direction code refers to the encoded field in the suspicious word that indicates the sampling direction of the intelligent sensing system. The sampling direction code is bound to the vehicle coordinates of the inspection robot and is used to restore the vector representation of the sampling direction in the vehicle coordinates. The process of extracting the three corner codes is as follows: according to the fixed length and bit width of the Euler angle quantization code, extract the first corner code, the second corner code, and the third corner code sequentially from the high bit to the low bit of the Euler angle quantization code. The extraction order of the three corner codes is consistent with the writing order. The process of restoring the vehicle body sampling vector is as follows: read the axis number and direction symbol in the vehicle body coordinates according to the sampling direction code, write the direction symbol into the corresponding axis, write the remaining axis to zero, and obtain the vehicle body sampling vector; S22. For each angle value, write the cosine of the angle value into the two coaxial holding positions of the rotation table, write the sine and negative sine of the angle value into the two opposite-axis reversing positions of the rotation table, write one into the rotation axis holding position of the rotation table, write zero into the remaining positions of the rotation table, and multiply the three rotation tables in order according to the three angle codes to generate a rotation matrix. The preserved position refers to the position of two elements located on the main diagonal of the same plane in the rotation table. After the coordinate rotation, the cosine projection of the corresponding plane component is retained. The reversal position refers to the position of two elements located on the anti-diagonal line of the same plane in the rotation table. After the coordinate rotation, the two element positions are written into the sine projection and the reverse sine projection of the corresponding plane components. The remaining bits refer to the positions of elements in the rotation table other than the two hold bits, the two reversal bits, and the rotation axis hold bits. The remaining bits are written as zero to prevent non-rotational plane components from participating in the conversion.

[0007] In a preferred embodiment, S2 further includes: S23. Obtain the station observation vector by rotating the vehicle body sampling vector to the left using the rotation matrix, and read the component test end coordinates registered by the electrical performance testing device using the electrical component index of the suspicious word. The electrical performance testing device refers to the device used to test the electrical parameters of electrical components and register the coordinates of the test end of the components. The coordinates of the test end of the components are used to generate the component direction vector and participate in the calculation of the verification action word. S24. Subtract the current position of the inspection robot from the coordinates of the component test end to obtain the component direction vector. Subtract the component direction vector from the station observation vector to obtain the source verification difference vector. Write the source verification action word according to the axis sign of the source verification difference vector.

[0008] In a preferred embodiment, S3 includes: S31. Based on the source action word, read the source difference vector and source action number. The edge computing unit recursively generates candidate actions according to the robot control resolution. For each candidate action generated, the source difference vector is obtained by subtracting the source difference vector from the sampled vector after the candidate action. The residual vector is then multiplied by itself and written into the action value. The source verification action number refers to the sequential identifier assigned by the edge computing unit to the source verification action word. The source verification action number is used to associate the execution action word, action completion bit, component electrical response, cable echo, and re-sampling word. The process of recursively generating candidate actions is as follows: The edge computing unit starts from the current vehicle position and the current robot Euler angle, and gradually increases the single-axis displacement pulse or single-axis rotation pulse according to the robot control resolution. Each time a pulse is added, the vehicle position, the robot Euler angle and the sampled vector after the action are calculated, and the corresponding pulse group is written as a candidate action. S32. For candidate actions corresponding to action value, the edge computing unit calculates the difference sign between the vehicle body position after the action and the boundary of the inspection channel, and calculates the difference sign between the test contact and the access point of the electrical performance test device after the action. If any difference sign exceeds the boundary, the candidate action is deleted. The candidate actions are retained and the first item is written as the execution action word in ascending order of action value. The test contact point refers to the measurement position on the vehicle body end of the inspection robot after the candidate action, which is used to connect to the electrical performance testing device. The test contact point is calculated by the vehicle body position after the action, the Euler angle of the robot after the action, and the measurement offset at the vehicle body end. The process of calculating the sign of the difference between the vehicle body position after the action and the inspection channel boundary is as follows: First, unify the vehicle body position after the action and the inspection channel boundary to the station coordinates. Then, generate the boundary distance value according to the directed distance from the outer contour point of the vehicle body after the action to the boundary line of the inspection channel. Write the boundary crossing symbol when the boundary distance value is negative, and write the non-boundary crossing symbol when the boundary distance value is non-negative. The process of calculating the sign of the difference between the test contact and the access point of the electrical performance testing device after the action is as follows: First, unify the test contact and the access point of the electrical performance testing device to the station coordinates. Then, subtract the coordinates of the access point of the electrical performance testing device from the coordinates of the test contact to generate the access difference vector. Perform self-multiplication and summation on the access difference vector to generate the access distance value. Finally, generate the access difference sign according to the sign of the difference between the access distance value and the allowable access distance of the electrical performance testing device.

[0009] In a preferred embodiment, S3 further includes: S33. Drive the inspection robot to change the sampling posture according to the execution action word. Recalculate the sampling vector with real-time Euler angles in each control cycle, and obtain the cycle residual vector by subtracting the source difference vector from the recalculated sampling vector. Write the sign of the difference between the sum of self-multiplication of the current cycle residual vector and the sum of self-multiplication of the previous cycle residual vector when the value changes from negative to zero. S34. At the action completion position, the edge computing unit triggers the electrical performance testing device to obtain the component electrical response with the electrical component index of the suspicious word, and triggers the electrical fault detection device to obtain the cable echo with the cable segment index of the suspicious word. Then, the component electrical response is connected to the cable echo and written as the re-collection word according to the source verification action number.

[0010] In a preferred embodiment, S4 includes: S41. Read the suspicious word and the resampled word according to the sampling order of the source sequence. The neural encoder generates an input difference by subtracting the electrical measurement value of the previous word from the electrical measurement value of the current word. Multiply the input difference by the encoding weight, add the previous response code multiplied by the recursive weight, and write the summation value as the response code according to the encoding bit width. The neural encoder refers to the recursive encoder deployed in the edge computing unit. The recursive encoder takes the input difference of the current word and the previous response code as input, performs multiplication, addition and rounding according to fixed encoding weights, recursive weights and encoding bit width, and outputs the response code of the current word. Electrical measurement values ​​refer to the component electrical response values ​​obtained by the electrical performance testing device in the suspected word or re-sampled word. The electrical measurement values ​​are used to participate in the calculation of adjacent differences according to the sampling order, and are used to characterize the changes in electrical performance response before and after the action. The coding weight refers to the fixed-point weight multiplied by the input difference in the recursive encoder. The coding weight is obtained by the edge computing unit reading the change in electrical measurement value of the same type of component and the confirmation result in the most recent manually confirmed fault diagnosis record, and dividing the number of times the sign of the change is consistent with the confirmation result by the total number of records of the same type of component. The response code is an integer code obtained by the recursive encoder after performing fixed-length encoding on the electrical changes of the current word. The response code is used to construct the grid value and serves as the alignment input for the neural dynamic time warping algorithm. S42. Establish grid points by using the response code of the suspicious word as the row code and the response code of the repeated word as the column code. First, calculate the absolute value of the difference between the row code and the column code. Then, combine the sum of the bitwise XOR values ​​of the suspicious word object index and the repeated word object index into the absolute value of the difference to generate the grid point value.

[0011] In a preferred embodiment, S4 further includes: S43. Recursively deduce the path table from the start position of the grid point to the end position of the grid point. For each grid point, read the cumulative value of the three predecessor grid points, take the first item in ascending order of value, add it to the current grid point value, and write it into the current cumulative value. Write the predecessor direction corresponding to the first item into the backtracking position. The predecessor grid point refers to the calculated grid point in the path table that can reach the current grid point, including the grid point in the same column of the previous row, the grid point in the previous column of the same row, and the grid point in the previous column of the previous row. The predecessor direction refers to the path direction mark from the predecessor grid point to the current grid point. The path direction mark is used to read the grid point in reverse order by backtracking after the path table is written to the end of the grid point. S44. After the path table is written to the end of the grid point, the grid point is read in reverse order from the end of the grid point according to the backtracking bit. When a grid point with a non-zero XOR sum of the object index is read, it is skipped. When a grid point with a zero XOR sum of the object index is read, it is written to the chain bit. The verification source alignment chain is generated in reverse order of the chain bit.

[0012] In a preferred embodiment, S5 includes: S51. Read the suspicious word and the repeated word one by one along the verification source alignment chain. Subtract the electrical measurement response value of the suspicious word from the electrical measurement response value of the repeated word, and generate the response difference direction according to the sign of the difference. The electrical response value refers to the digital code obtained by the electrical performance testing device from the suspected word or the re-acquisition word after conversion by the same dimension calibration table. The conversion is based on the reference unit, zero offset and quantization step size registered by the electrical performance testing device. The original component electrical response is subtracted from the zero offset and then divided by the quantization step size and rounded to the nearest integer, so that the suspected word and the re-acquisition word can be subtracted on the same numerical scale. The difference sign is used to indicate the direction of electrical measurement change before and after the source verification action. S52. Based on the source action word, read the source direction, multiply the response difference direction with the source direction bit by bit, delete the current chain bit when the product is negative, and write the current chain bit into the candidate chain when the product is non-negative.

[0013] In a preferred embodiment, S5 further includes: S53. Read the electrical performance test value and fault detection value under the same object index for the candidate chain, multiply the sign of the difference between the electrical performance test value and the sign of the difference between the fault detection value, delete the current chain position when the product is negative, and write the current chain position into the verification chain when the product is non-negative. S54. Read the connection path corresponding to the object index in the verification chain along the electrical connection table, XOR the endpoint index of the connection path with the fault semantic index bit by bit, output the corresponding fault code when the XOR sum is zero, delete the current chain bit when the XOR sum is non-zero, and generate the inspection robot fault diagnosis result according to the retained fault code. The electrical connection table refers to the topology index table generated from the electrical equipment ledger and line connection records. The table uses the object index as the read key and records the cable segment index, connection path and endpoint index corresponding to the object index. It is used to check whether the suspicious words in the check chain are located in the same electrical connection relationship. Fault semantics refers to the electrical meaning index corresponding to the fault code. The electrical meaning index writes the object index, cable segment index, and endpoint index that the fault code should apply to into the same encoding field, which is used to perform bitwise XOR with the endpoint index of the connection path to determine whether the reserved chain bits can form an electrically interpretable fault diagnosis result.

[0014] The technical effects and advantages of this invention are as follows: 1. This solution establishes a chain of evidence before and after an action by using suspicious words, source verification action words, and re-collection words, so that abnormal evidence is retained after being verified by the re-collection response, which relatively reduces the risk of multimodal anomalies being incorrectly bound to the same fault object; 2. Write the Euler angle quantization code, electrical component index, and cable segment index into the sampling word consecutively to keep the attitude, component, and line affiliation in the same sampling order, thereby reducing diagnostic bias caused by cross-object data retrieval. 3. The station observation vector is obtained by Euler angle coordinate transformation and the source verification action word is calculated with the component direction, so that the re-sampling action can be deduced from the source of doubt, which relatively improves the re-verification of the limited dwell time; 4. Obtain the component's electrical response and cable echo at the action completion position, so that the re-sampling results simultaneously cover the component and line responses, which helps to distinguish between electrical component abnormalities and cable section abnormalities; 5. Perform neural dynamic time warping on suspicious words and repeated sampling words to generate a source alignment chain based on the response code and object index, which can alleviate the response mismatch caused by the difference in sampling rhythm during mobile inspection. Attached Figure Description

[0015] Figure 1 This is a roadmap for the multimodal source detection and diagnostic technology of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Refer to the instruction manual appendix Figure 1 The present invention provides a fault diagnosis method for power inspection robots based on multimodal spatiotemporal alignment and three-layer fusion inference, comprising: S1. Obtain the inspection package read by the edge computing unit of the inspection robot from the intelligent sensing system, generate sampling words according to the sampling order, the sampling words are composed of Euler angle quantization code followed by electrical component index and then cable segment index, and reverse the position of the electrical measurement difference sign of adjacent sampling words to write the suspicious point word. To address the formation of suspicious words, S1 first solidifies the robot posture, electrical component affiliation, line section affiliation, and electrical measurement readings under the same sampling sequence into the same data unit, and then uses the changing direction of adjacent sampling sequences to determine the electrical measurement reversal position that needs to be verified. When S2 reads suspicious words subsequently, the source of changes in robot Euler angles, part numbers, line segment numbers, and electrical measurements all originate from the same sampling sequence, reducing the risk of data from different objects being mixed into the same diagnostic chain from the data entry point; this implementation process includes the following steps: As the inspection robot moves along the substation passage, the intelligent sensing system generates on-site reading results at the same trigger moment, and the edge computing unit encapsulates the on-site reading results into an inspection package. The inspection package uses the incremental sampling order as an index to write the robot Euler angles output by the robot posture sensing unit, the part number mapped from the electrical equipment ledger, the line segment number mapped from the line connection record, the electrical performance test value output by the electrical performance testing device, and the fault detection value output by the electrical fault detection device into the same field record frame. When executing S11, the edge computing unit converts the robot's Euler angles into angle units. Negative angles are first converted to non-negative angles by adding 360 degrees, then divided by the angle resolution and the integer quotient is used to obtain the Euler angle quantization code. The angle resolution comes from the output resolution of the robot's posture sensing unit or the posture configuration table registered by the edge computing unit. The width of the Euler angle quantization code is determined by the upper limit of the encoding after dividing 360 degrees by the angle resolution; the width of the component number is determined by the number of components in the electrical equipment ledger; and the width of the line segment number is determined by the number of cable segments in the line connection record. When a field is less than its width, pad it with zeros in the high bits; when a field exceeds its width, stop writing the sampling word and output the index to reallocate the record. When the field is complete, the edge computing unit uses the sampling order as the write key to write the Euler angle quantization code, part number and line segment number into the sampling word in a fixed length and bit width, and reads it when encoding the power measurement direction; After the sampling word is generated, the edge computing unit enters the direction encoding process of S12: according to the component number in the current sampling word, read the electrical performance test value of the same component number under the current sampling order and the previous sampling order, and subtract the previous electrical performance test value from the current electrical performance test value. If the difference is positive, write 1; if the difference is negative, write negative 1; if the difference is zero, write zero, and generate the component difference sign. At the same time, the edge computing unit reads the fault detection value of the same line segment number in the current sampling order and the previous sampling order according to the line segment number in the current sampling word, and subtracts the previous fault detection value from the current fault detection value. If the difference is positive, write 1; if the difference is negative, write negative 1; if the difference is zero, write zero, and generate the line segment difference sign. When electrical performance test values ​​are missing, write zero for the component difference sign and record the missing component position; when fault detection values ​​are missing, write zero for the line segment difference sign and record the missing line segment position. After the component difference symbol and the line segment difference symbol are generated, the edge computing unit multiplies the two to obtain the electrical measurement symbol bit of the current sampling sequence, which is read when a suspicious point is triggered. During the movement of the inspection robot, there are reflections, distance changes and line response lags. In S13, the edge computing unit filters the electrical test direction reversal position according to the same part number and the same line segment number: the edge computing unit reads the current electrical test symbol bit according to the sampling order, and retrieves the previous non-zero electrical test symbol bit that is closest to the current sampling order under the same part number and the same line segment number. When the current electrical measurement symbol bit is zero, the edge computing unit retains the current sampling word and continues to read the next sampling sequence bit; When the product of the current electrical measurement sign bit and the previous non-zero electrical measurement sign bit is negative, the edge computing unit determines that there is a joint reversal of the direction of the component electrical response and the line fault response in the current sampling sequence, and writes the sampling word corresponding to the current sampling sequence as the suspicious word. When there is no previous non-zero electrical measurement sign bit, the edge computing unit writes the current sampling sequence into the starting record and does not generate a suspicious word. When writing suspicious words, the sampling sequence, Euler angle quantization code, part number, line segment number, electrical test symbol bit and inspection package index are simultaneously retained for S2 to read and restore the robot's Euler angle and sampling direction; Through the above processing, S1 generates a suspicious word containing the attitude source, object attribution, line attribution, and electrical measurement reversal basis; when S2 generates the source verification action word based on the suspicious word, it can read back the robot's Euler angle along the same sampling sequence, and read back the source of electrical measurement change along the same part number and the same line segment number. In practical applications: When the inspection robot passes through the passage in front of the switch cabinet, the edge computing unit reads the electrical performance test value of the same component number changing from rising to falling at the fifteenth sampling position, and at the same line segment number changing from falling to rising. The electrical test symbol bit after multiplying the two direction symbols is reversed relative to the previous non-zero electrical test symbol bit. The edge computing unit then writes the sampling word of the fifteenth sampling position as the suspicious word, so that S2 can continue to calculate the source verification action word.

[0018] S2. Based on the suspicious words, restore the robot's Euler angles, generate the rotation matrix from the vehicle coordinates to the station coordinates through the Euler angle coordinate transformation algorithm, convert the sampling direction of the suspicious words into the station observation vector, and then subtract the direction corresponding to the electrical component index to generate the source verification action word. After the suspicious word enters S2, the edge computing unit performs calculations around "whether the current sampling direction points to the actual electrical test object": First, it recovers the vehicle posture and sampling direction from the suspicious word, then converts the sampling direction in the vehicle coordinates to the station coordinates, and then reads the component test end coordinates registered by the electrical performance testing device, using the difference between the observation direction and the component direction to generate the source verification action word; the specific execution is as follows: The Euler angle quantization code and sampling direction code in the suspicious word are responsible for attitude restoration and sampling direction restoration, respectively. When the edge computing unit executes S21, it first reads the sampling direction code in the extended field of the suspicious word. The sampling direction code is written by the sampling channel direction table of the intelligent sensing system. The field contains the axis number and direction symbol of the vehicle coordinate. Then, according to the fixed length and bit width of the Euler angle quantization code, the first angle code, the second angle code, and the third angle code are extracted from the high bit to the low bit. The first angle code corresponds to the vertical axis rotation within the station, the second angle code corresponds to the horizontal axis rotation of the car body, and the third angle code corresponds to the longitudinal axis rotation of the car body. The three angle codes are then multiplied by the angle resolution to obtain the three angle values. During the sampling direction restoration, the edge computing unit writes the direction symbol in the sampling direction code into the corresponding axis position and writes zero into the remaining axis positions to generate a three-row, one-column vehicle body sampling vector. When the Euler angle quantization code is missing, an attitude missing record is written; when the sampling direction code is missing, a direction missing record is written, and the generation of the source verification action word is paused. The rotation matrix is ​​generated under the constraint of the axis order of the three angle values. In S22, the edge computing unit constructs a three-row, three-column rotation table for each angle value. The rows and columns of the rotation table are registered in the order of the three axes of the station coordinates. For any angle value, the edge calculation unit writes the cosine of the angle value into two holding positions of the rotation plane, writes the sine and negative sine of the angle value into two reversing positions of the rotation plane, writes 1 into the rotation axis holding position, and writes zero into the remaining positions; The holding position is the position of two elements on the main diagonal of the rotation plane, the reversing position is the position of two elements on the anti-diagonal of the rotation plane, and the remaining positions are the positions of elements other than the holding position, the reversing position, and the holding position of the rotation axis. The three rotation tables are multiplied in the order of writing the first, second, and third subcodes to obtain the rotation matrix from the vehicle coordinates to the station coordinates. If there are empty spaces in the rotating table, fill them with zeros and check the number of digits in the three rows and three columns. If the number of digits is inconsistent, write an error record to the rotating table. After completing the attitude conversion, S23 puts the sampling direction and the position of the electrical test object into the same in-station coordinates for calculation. The edge calculation unit multiplies the three-row-one-column vehicle body sampling vector by the rotation matrix to obtain the three-row-one-column in-station observation vector. Next, the coordinates of the component test end registered by the electrical performance testing device are read from the electrical component index in the question mark. The coordinates of the component test end are in-station coordinates and correspond one-to-one with the electrical component index. In this embodiment, the electrical performance testing device is responsible for testing the electrical parameters of electrical components and registering the coordinates of the component test ends. The coordinates of the component test ends are used for calculating the component's direction vector. When the electrical component index does not match the registration form, the edge computing unit writes the missing record to the test end and transfers the suspicious words to the manual review queue; The source verification action word is calculated from the direction difference in S24. The edge computing unit reads the current position of the inspection robot and subtracts the current position of the inspection robot from the coordinates of the component test end to obtain the component coordinate difference vector. To unify the units, the edge computing unit performs self-multiplication and summation on the component coordinate difference vector and takes the square root to obtain the component distance value. Then, it divides the component coordinate difference vector by the component distance value to generate a normalized component direction vector. Then, the source difference vector is obtained by subtracting the component direction vector from the station observation vector. The axis sign of the source difference vector is read in the order of the three axes of the station coordinate system: positive value is written as 1, negative value is written as negative 1, and zero value is written as zero. The source verification action word is written with the source verification action number, the suspicious word index, the source verification difference vector, and the source direction. The source direction is composed of the axis sign of the source verification difference vector, which is used by S3 to generate candidate actions and by S5 to perform direction verification. When the component distance value is zero, the component direction vector is written as the zero vector, and the current position coincidence is recorded. The above processing converts the attitude encoding, sampling direction, and component test end coordinates in the suspicious words into executable source verification action words, enabling S3 to drive the inspection robot to re-sample based on the direction difference in the station coordinates; In practical applications: When the inspection robot passes in front of the switch cabinet, the edge computing unit recovers the three-segment angle values ​​and the vehicle body sampling vector from the suspicious words, and obtains the station observation vector through the rotation matrix, and then reads the coordinates of the component test terminal corresponding to the wiring terminal; If the observation vector within the station is biased towards the left side of the cabinet while the component direction vector points towards the middle terminal of the cabinet, the horizontal axis symbol of the source difference vector is written into the source action word. Subsequently, S3 adjusts the sampling posture and obtains the re-sampling word accordingly.

[0019] S3. Drive the inspection robot to change the sampling posture according to the source verification action word. At the action completion position, the electrical performance testing device obtains the component's electrical response, the electrical fault detection device obtains the cable echo, and writes it as the re-sampling word according to the source verification action number. S3 is used to convert the source verification action word into the field re-collection action, and write the electrical test result after the action is completed into the re-collection word; The edge computing unit first recursively derives candidate actions based on the source difference vector, then filters the action words to be executed using the inspection channel boundary, the access location of the electrical performance testing device, and the action cost. Subsequently, within the control cycle, it determines whether the action has reached the source location based on real-time Euler angles, and finally triggers the electrical performance testing device and the electrical fault detection device to obtain re-collected data. This implementation process includes the following steps: When the source verification action word enters the action planning, the edge computing unit reads the source verification difference vector and the source verification action number in S31. The source verification action number is the sequence identifier assigned by the edge computing unit to the source verification action word, which is used to keep the execution action word, action completion bit, component electrical response, cable echo and re-collection word under the same source verification task. The candidate actions are recursively derived from the current vehicle position and the current robot Euler angle. The recursion unit is given by the robot control resolution, which comes from the single-axis displacement pulse length and single-axis rotation pulse angle registered by the robot controller. Each time the edge computing unit adds a single-axis displacement pulse or a single-axis rotation pulse, it calculates the vehicle position after the action, the robot's Euler angles after the action, and the sampled vector after the action, and writes the current pulse group as a candidate action; The post-action sampling vector is obtained by converting the robot's Euler angles after the action using the same Euler angle coordinate transformation rule as S2, ensuring that the action cost is calculated only within the vector dimension. For each candidate action generated, the edge computing unit subtracts the source difference vector from the action post-sampling vector to obtain the residual vector, and then multiplies each axis component of the residual vector and sums them to obtain the action cost. The candidate action is pushed to the mechanical attitude travel boundary, the inspection channel boundary, or the upper limit of the single source verification action registered by the controller, and stops when no candidate action is generated. If no candidate action is generated, an action unreachable record is written. Candidate actions are selected for execution only after passing the safety boundary and access boundary. In S32, the edge computing unit first unifies the vehicle body position and inspection channel boundary after the action to the station coordinates, and calculates the vehicle body outline point after the action based on the vehicle body position after the action, the robot Euler angle after the action, and the vehicle body outline size. After each action, the directed distance from the outer contour point of the vehicle body to the boundary line of the inspection channel is calculated in the unit of coordinate length within the station. When any directed distance is negative, a boundary crossing symbol is written; when all directed distances are non-negative, a boundary not crossing symbol is written. The test contact point is obtained by multiplying the rotation matrix generated by the Euler angles of the robot after the action by the offset measured at the left end of the vehicle body, and then adding the position of the vehicle body after the action. The offset measured at the vehicle body end comes from the calibration table installed at the measuring end of the inspection robot. The edge computing unit unifies the test contact and the electrical performance testing device access point to the station coordinates. It generates an access difference vector by subtracting the electrical performance testing device access point coordinates from the test contact coordinates. It performs self-multiplication and summation on the access difference vector to generate the access distance value. Then, it subtracts the allowable access distance of the electrical performance testing device from the access distance value to generate the access difference sign. The allowable access distance comes from the mechanical tolerance of the contact end or the effective non-contact measurement distance registered by the electrical performance testing device. When a boundary crossing symbol appears or the access difference symbol is positive, the corresponding candidate action is deleted. The candidate actions are retained and sorted in ascending order of action value. When the action values ​​are the same, the first item is selected in the order of displacement pulse number, rotation pulse number, and source verification action number, and written as the execution action word. After the action word is issued, the inspection robot changes the sampling posture according to the control cycle in S33. The edge computing unit reads the real-time Euler angles in each control cycle and recalculates the sampling vector according to the same Euler angle coordinate transformation rule as in S2. The periodic residual vector is obtained by subtracting the source difference vector from the recalculated sampling vector. The periodic residual value is obtained by multiplying and summing the components of each axis of the periodic residual vector. To avoid relying on thresholds that cannot be implemented, the periodic residual value is quantized according to the minimum vector change corresponding to the robot control resolution. The periodic residual change value is obtained by subtracting the previous quantized periodic residual value from the current quantized periodic residual value. If the periodic residual change value is negative, write negative one; if the periodic residual change value is zero, write zero; if the periodic residual change value is positive, write one. When the sign of the current cycle residual change is zero and the sign of the previous cycle residual change is negative, the edge computing unit writes the action completion bit; if the control cycle reaches the last bit of the pulse group registered by the action word and the action completion bit is not written, the edge computing unit stops the action and writes the action incomplete record, and the suspicious word is retained to the manual review queue. After the action completion bit is written, S34 triggers the acquisition of re-acquisition data and the encapsulation of re-acquisition word; the edge computing unit triggers the electrical performance testing device with the electrical component index of the suspicious word, the electrical performance testing device acquires the component electrical response at the position corresponding to the action completion bit, and writes the component electrical response into the edge computing unit according to the reference unit registered by the electrical performance testing device; The edge computing unit triggers the electrical fault detection device with the cable segment index of the suspicious word. The electrical fault detection device obtains the cable echo at the same action completion position and writes the cable echo into the edge computing unit according to the echo unit registered by the fault detection device. The repeated sampling characters are written sequentially according to the source verification action number, suspicious character index, component electrical response, cable echo and action completion bit. The bit width of the source verification action number comes from the total number of source verification actions in the current inspection task. The bit width of the component electrical response comes from the range and quantization step size of the electrical performance testing device. The bit width of the cable echo comes from the range and sampling resolution of the electrical fault detection device. When a component's electrical response is missing, a missing bit is written to the component's re-acquisition bit; when a cable echo is missing, a missing bit is written to the cable's re-acquisition bit. If any missing bit exists, the re-acquisition word is still saved, but it participates in the alignment cost calculation in S4 based on the missing bit. The above processing transforms the source verification action word into a re-collection action constrained by the channel boundary, test access conditions, and action residual, and binds the component electrical response and cable echo after the action to the same source verification action number, so that S4 can construct source verification sequence pairs; In practical applications: When the inspection robot faces the switch cabinet terminal, the source difference vector indicates that the sampling direction needs to be adjusted towards the center of the cabinet. After the edge computing unit recursively deleting multiple candidate actions that would cross the channel boundary or fail to connect to the test terminal, the action word with the highest action value is selected. After the robot completes its posture adjustment, the electrical performance testing device reads the electrical response at the terminal, the electrical fault detection device reads the echo of the corresponding cable segment, and the edge computing unit writes the two re-collection results along with the source verification action number into the re-collection word.

[0020] S4. Perform the neural dynamic time warping algorithm on the source verification sequence pair composed of the suspicious word and the re-collected word. The neural encoder generates response codes, writes the absolute value of the difference between response codes at the same index position into the grid cost table, and writes the first item into the path table in ascending order of the cumulative cost of the three predecessor paths. Backtrack to generate the source verification alignment chain. After the source verification sequence reaches the edge computing unit, S4 rewrites the suspicious word before the action and the re-acquisition word after the action into the same response code space, and then uses the response code difference and object index consistency to jointly constrain the alignment path, so that S5 can judge the change of electrical measurement response of the same electrical object before and after the source verification action along the source verification alignment chain; this implementation process includes the following steps: The response code originates from the sequential change of electrical measurement values. In S41, the edge computing unit reads the suspicious word sequence and the re-sampled word sequence according to the sampling sequence. The suspicious word sequence and the re-sampled word sequence share the neural encoder deployed in the edge computing unit. The neural encoder is a recursive encoder. Its input is the input difference of the current word and the previous response code, and its output is the response code of the current word. The electrical test value is taken from the component electrical response value in the suspicious word or the repeated word. The previous electrical test value of the first word in the sequence is taken from the electrical test value of the first word. The previous response code of the first word in the sequence is written as zero. The input difference is generated by subtracting the previous word electrical test value from the current word electrical test value. The encoding weight is multiplied by the input difference, the recursive weight is multiplied by the previous response code, and the two products are added together and rounded down according to the encoding bit width to form the response code. The same type of component is determined by the equipment type code in the electrical equipment ledger. The coding weight is obtained by the edge computing unit reading the fault diagnosis record of the same type of component that has been manually confirmed. If the confirmation result is that the fault is established, write 1; if the fault is not established, write negative 1. When the sign of the change in electrical measurement value is multiplied by the confirmation result and is positive, it is counted as the number of consistency. The coding weight is the number of consistency divided by the total number of records of the same type of component. When the total number of records of the same type of component is zero, the coding weight is written as 1. The recursive weight is calculated by dividing the number of times adjacent response codes have the same sign in the records of the same type of component by the total number of records of the same type of component. When the total number of records of the same type of component is zero, the recursive weight is written as zero. The encoding bit width is determined by the number of response code storage bits registered by the edge computing unit. When the summation value is higher than the upper limit of the encoding bit width, the upper limit is written; when it is lower than the lower limit of the encoding bit width, the lower limit is written. The grid value records both the response difference and the object mismatch cost in S42. The edge computing unit uses the suspicious word response code as the row code and the re-sampling word response code as the column code, and establishes grids according to the row code order and column code order. For each grid point, the absolute value of the difference between the row code and the column code is first calculated, and then the object index of the suspicious word and the object index of the repeated word are read. The object index is formed by the electrical component index and the cable segment index. The edge computing unit performs a bitwise XOR operation on the two object indices and accumulates the XOR bits to obtain the object XOR sum. Then, it adds the absolute value of the response code difference to the object XOR sum and writes it as the grid value of that grid point. When the response code is missing, the absolute value of the response code difference is written to the upper bound of the encoding bit width. When the object index is missing, the XOR sum of the object is written to the object index bit width. This allows the missing data to still enter the path table, but retains the missing penalty in the cumulative cost. When the path table advances along the grid cost table, S43 adopts the recursive rule of neural dynamic time warping: the cumulative cost of the grid starting position is directly written as the cost of that grid point. The first row of grid points reads the grid point in the previous column of the same row as the predecessor grid point. The first column of grid points reads the grid point in the same column of the previous row as the predecessor grid point. The remaining grid points read the grid point in the same column of the previous row, the grid point in the previous column of the same row, and the grid point in the previous column of the previous row. The edge computing unit sorts the cumulative cost of the readable predecessor grid points in ascending order. When the values ​​are the same, the first item is taken according to the fixed order of the previous row and column, the same column of the previous row, and the previous column of the same row. The first item is then added to the current grid point cost to obtain the current cumulative cost, and the predecessor direction corresponding to the first item is written to the backtracking bit. The preceding direction is the direction marker for the current grid point to continue from the previous grid point, which is used to read the path backward from the last grid point in the subsequent process. Backtracking reads are initiated in S44 from the last grid position. The edge computing unit reads back grid by grid position in reverse order until it returns to the starting grid position. For each grid point read, the edge computing unit reads the XOR sum of the object corresponding to that grid point. If the XOR sum of the object is non-zero, the grid point is skipped; if the XOR sum of the object is zero, it is written to the chain bit. The chain bits are written to the suspicious word index, the repeated sampling word index, the grid value and the backtracking bit. After the reverse reading is completed, the edge computing unit will arrange the chain bits in reverse order to generate the source alignment chain. When an empty backtracking bit appears in the path table, the edge computing unit retains the already written chain bit and records the path interruption bit. When the source verification alignment chain is empty, the source verification sequence pair is transferred to the manual review queue. With the above processing, the suspected words before the action and the re-acquired words after the action form a source alignment chain under the same electrical component index, the same cable segment index and similar response code changes. S5 can calculate the response difference direction along the chain and delete the incorrectly bound suspected words. In practical applications: After the inspection robot completes the source verification action on the terminal, the edge computing unit encodes the component electrical response changes in the suspicious word and the component electrical response changes in the re-collection word into response codes respectively; the XOR sum of the grid objects corresponding to the same terminal and the same cable segment is zero, and the grid is written into the chain position when the path table is traced back, finally forming the source verification alignment chain for S5 to check.

[0021] S5. Calculate the response difference direction before and after the verification action along the verification source alignment chain. Multiply the response difference direction with the source direction of the verification action word bit by bit. Delete suspicious words with negative products from the verification source alignment chain. The retained suspicious words are checked for consistency between electrical detection rules, electrical connection and fault semantics. Output the inspection robot fault diagnosis result. The source alignment chain received by S5 has already given the correspondence between the suspicious words and the re-collected words. This implementation process continues to complete three types of verification along the chain position: First, the electrical response of the component before and after the action is converted to the same dimension and the direction of the response difference is generated. Then, the chain position opposite to the source action is deleted using the source direction. Subsequently, the verification chain is screened out using the consistency of the direction of the electrical performance test value and the fault detection value. Finally, the inspection robot fault diagnosis result is output by combining the electrical connection table and the fault semantic index. This implementation process includes the following steps: Each link in the source alignment chain simultaneously stores the suspicious word index and the re-sampling word index. When the edge computing unit executes S51, it reads the original component electrical response in the suspicious word and the original component electrical response in the re-sampling word along the link and calls the same dimension calibration table registered by the electrical performance testing device. The same dimension calibration table is written with the reference unit, zero offset and quantization step size. The edge computing unit first converts the original component electrical response to the reference unit, then subtracts the zero offset from the converted original component electrical response, divides the difference by the quantization step size and rounds it to obtain the electrical response value of the suspicious word and the electrical response value of the re-acquisition word respectively. The response difference is obtained by subtracting the electrical response value of the suspicious character from the electrical response value of the repeated character. A positive response difference is written as 1, a negative response difference as negative 1, and a zero response difference as zero. To ensure that subsequent bit-by-bit multiplication is executable, the edge computing unit generates the response difference direction according to the bit width of the source direction in the verification action word. The non-zero axis in the source direction is written with the response difference sign, and the zero axis in the source direction is written with zero. When the dimensional calibration table is missing, the edge computing unit writes the current chain position into the calibration missing record and stops writing the current chain position into the candidate chain; The source direction comes from the source difference vector axis symbol written when S2 generates the source verification action word. The edge computing unit reads the source verification action word corresponding to the current chain position in S52 and obtains the source direction according to the three axes of the station coordinates. Both the response difference direction and the source direction are three-bit symbol codes. Positive values ​​are written as 1, negative values ​​as -1, and zero values ​​as zero. The edge computing unit multiplies the two symbol codes bit by bit to obtain a three-bit product code. When any non-zero bit in the three-bit product code is negative, it indicates that the direction of electrical measurement change after the action is opposite to the direction of the axis position source, and the edge calculation unit deletes the current chain position; When all three product bits are zero or all non-zero bits are positive, the current chain bit does not form reverse evidence, and the edge computing unit writes the current chain bit into the candidate chain; when the source direction is missing, the current chain bit is written into the source missing record and retained in the manual review queue. The candidate chain continues to undergo joint directional verification of electrical performance test values ​​and fault detection values. In S53, the edge computing unit reads the current chain position and the previous chain position according to the candidate chain order, and limits the reading range with the object index. The object index is formed by the electrical component index and the cable segment index. Under the same electrical component index, the edge computing unit subtracts the electrical performance test value of the previous chain bit from the current chain bit's electrical performance test value. If the difference is positive, write 1; if it is negative, write negative 1; if it is zero, write zero, and generate the electrical performance test value difference sign. Under the same cable segment index, the edge computing unit subtracts the fault detection value of the previous chain position from the fault detection value of the current chain position. If the difference is positive, write 1; if it is negative, write negative 1; if it is zero, write zero, and generate the sign of the fault detection value difference. When the first link of a candidate chain has no preceding link, both the sign of the electrical performance test value difference and the sign of the fault detection value difference should be written as zero; When the product of the two difference signs is negative, the change in the component's electrical response at the current link position conflicts with the change in the line fault response, and the edge computing unit deletes the current link position; when the product is zero or positive, the current link position is written into the verification link. When electrical performance test values ​​or fault detection values ​​are missing, the corresponding difference sign is written as zero, and the missing bit is synchronously written into the verification chain; The output of the verification chain needs to be applied to electrical connection relationships and fault semantic relationships. In S54, the edge computing unit reads the electrical connection table using the object index of the verification chain. The electrical connection table is generated from the electrical equipment ledger and line connection records. The table uses the object index as the read key to write the electrical component index, cable segment index, connection path number, start endpoint index, end endpoint index, and intermediate connection bit. The edge computing unit reads the endpoint index along the connection path and reads the fault semantic index with the candidate fault code as the key. The fault semantic index writes the electrical component index, cable segment index and endpoint index that the fault code should apply into the same encoding field. The endpoint index of the connection path is XORed bit by bit with the fault semantic index. When the XOR sum is zero, the edge computing unit writes the candidate fault code into the fault code table. When the XOR sum is non-zero, the current chain bit is deleted. After all the verification chains are read, the edge computing unit generates the inspection robot's fault diagnosis results according to the fault code table. The fault diagnosis results are written with the fault code, the corresponding electrical component index, the corresponding cable segment index, the corresponding chain position, and the source verification action number. When the electrical connection table does not hit the object index, the current link is written with a connection missing record and does not participate in the generation of the fault code table. After S5 is completed, the remaining chain positions have simultaneously met the verification requirements of action direction, electrical measurement direction, line response direction and electrical connection semantics. The fault diagnosis results of the inspection robot can be traced back to the suspicious words, re-collection words, source verification action numbers and endpoint indexes in the electrical connection table. In practical applications: After the inspection robot completes the source verification and re-sampling of the switch cabinet wiring terminals, the re-sampling electrical test response value of a certain chain position increases relative to the suspected point electrical test response value. The source direction is positive in the horizontal axis position, and no negative value appears after multiplying bit by bit. The direction of change of the fault detection value in the same cable segment is consistent with the direction of change of the electrical performance test value, and the chain position enters the verification chain; the electrical connection table shows that the terminal is connected to the end point of the cable segment, the XOR sum of the fault semantic index and the connection path end point index is zero, and the edge computing unit writes the corresponding fault code into the inspection robot's fault diagnosis result.

[0022] Among them Figure 1 It should be noted that the diagram shows the complete execution route of this solution from left to right and then from top to bottom. First, the intelligent sensing system reads the inspection package, which contains the sampling sequence, robot Euler angles, electrical performance test values, and fault detection values. Then, S1 writes the inspection package into sampling words and generates suspicious words based on the inversion of the electrical test symbol bits. S2 restores the Euler angles and sampling direction based on the suspicious words, obtains the in-station observation vector through Euler angle coordinate transformation, and generates the source verification action word in combination with the component direction. S3 recursively deduces candidate actions based on the source verification action word, selects the execution action word, and obtains the component electrical response and cable echo at the action completion position. S4 generates response codes through the neural encoder, and then uses the neural dynamic time warping recursive path table and backtracks to obtain the source verification alignment chain. S5 performs response direction, electrical connection, and fault semantic verification along the source verification alignment chain, and finally outputs the inspection robot fault diagnosis result. The solid arrows in the diagram represent the data and result transmission relationship of the main execution process, i.e., the inspection package, suspicious word, source verification action word, re-collection word, source verification alignment chain, and fault diagnosis result are generated sequentially; the dashed arrows represent the auxiliary input of external support data or constraints to the corresponding steps. For example, the electrical equipment ledger and line connection record support S1 to generate component number and line segment number, the electrical performance testing device registration form supports S2 to read the coordinates of the component test end, the inspection channel boundary and access allowable distance constraint S3 for candidate action screening, the manual confirmation record supports the coding weight and recursive weight values ​​in S4, and the electrical connection table and fault semantic index are used for the final consistency verification in S5.

[0023] Working Principle: This scheme first uses the edge computing unit of the inspection robot to read the inspection package from the intelligent sensing system. It writes the robot posture, electrical component affiliation, cable section affiliation, and electrical test readings into sampling words according to the sampling order. Then, it generates suspicious words by reversing the electrical test change direction of adjacent sampling words. Subsequently, it restores the robot's Euler angles and sampling direction based on the suspicious words. It converts the vehicle body sampling direction to in-station coordinates through Euler angle coordinate transformation. Then, it subtracts the component test end direction registered by the electrical performance testing device to obtain the source verification action word. The robot adjusts its sampling posture according to the source verification action word, and re-acquires the component electrical response and cable echo at the action completion position and writes it as a re-sampling word. Then, it aligns the suspicious words and re-sampling words through the neural dynamic time warping algorithm to form a source verification alignment chain. Finally, it compares the electrical test response direction before and after the action along the source verification alignment chain, and combines it with electrical testing rules, electrical connection relationships, and fault semantic verification to output the inspection robot fault diagnosis result. For example, when an inspection robot moves to inspect a switchgear in a substation, the edge computing unit discovers that the electrical performance test value corresponding to a certain terminal has reversed direction, and at the same time, the fault detection value of the corresponding cable section also shows a correlated change. Therefore, this sampling location is marked as a suspicious point. The robot calculates that the current observation direction deviates from the terminal test position based on the Euler angles and sampling direction at that time, and then adjusts its posture to resample towards the terminal. Subsequently, the electrical performance testing device reads the electrical response of the terminal component, and the electrical fault detection device reads the corresponding cable echo. After aligning the resampled results with the original suspicious point data, if the direction of the response change after the action is consistent with the direction of the source verification action, and the electrical connection table and fault semantics also point to the same terminal and cable section, then the corresponding fault diagnosis result is output. If the resampled response is opposite to the action direction or the connection relationship is inconsistent, the suspicious point is deleted to avoid misjudging reflections, distance changes, or interference from adjacent cabinets as real electrical faults.

[0024] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A fault diagnosis method for power line inspection robots based on multimodal spatiotemporal alignment and three-layer fusion inference, characterized in that, include: S1. Obtain the inspection package read by the edge computing unit of the inspection robot from the intelligent sensing system, generate sampling words according to the sampling order, the sampling words are composed of Euler angle quantization code followed by electrical component index and then cable segment index, and reverse the position of the electrical measurement difference sign of adjacent sampling words to write the suspicious point word. S2. Based on the suspicious words, restore the robot's Euler angles, generate the rotation matrix from the vehicle coordinates to the station coordinates through the Euler angle coordinate transformation algorithm, convert the sampling direction of the suspicious words into the station observation vector, and then subtract the direction corresponding to the electrical component index to generate the source verification action word. S3. Drive the inspection robot to change the sampling posture according to the source verification action word. At the action completion position, the electrical performance testing device obtains the component's electrical response, the electrical fault detection device obtains the cable echo, and writes it as the re-sampling word according to the source verification action number. S4. Perform the neural dynamic time warping algorithm on the source verification sequence pair composed of the suspicious word and the re-collected word. The neural encoder generates response codes, writes the absolute value of the difference between response codes at the same index position into the grid cost table, and writes the first item into the path table in ascending order of the cumulative cost of the three predecessor paths. Backtrack to generate the source verification alignment chain. S5. Calculate the response difference direction before and after the verification action along the verification source alignment chain. Multiply the response difference direction with the source direction of the verification action word bit by bit. Delete suspicious words with negative products from the verification source alignment chain. The retained suspicious words are checked for consistency between electrical detection rules, electrical connection continuity and fault semantics. Output the inspection robot fault diagnosis result.

2. The fault diagnosis method for power inspection robots based on multimodal spatiotemporal alignment and three-layer fusion inference according to claim 1, characterized in that: S1 includes: S11. After obtaining the inspection package, the edge computing unit uses the sampling order as the write key to generate the Euler angle quantization code by dividing the robot's Euler angle by the angle resolution, and writes the Euler angle quantization code, part number and line segment number into the sampling word in a fixed length and bit width. S12. Read the electrical performance test value based on the component number of the sampling word and generate the component difference symbol; read the fault detection value based on the line segment number of the sampling word and generate the line segment difference symbol; multiply the component difference symbol and the line segment difference symbol to generate the electrical test symbol bit. S13. Under the same sampling order, read the electrical measurement symbol bit. When the product of the current electrical measurement symbol bit and the previous electrical measurement symbol bit is negative, the edge computing unit writes the sampling word corresponding to the current sampling order as the doubt word.

3. The fault diagnosis method for power inspection robots based on multimodal spatiotemporal alignment and three-layer fusion inference according to claim 2, characterized in that: S2 includes: S21. Based on the suspicious words, read the Euler angle quantization code and sampling direction code, extract three segments of angle code according to the writing order of the Euler angle quantization code, multiply each segment of angle code by the angle resolution to obtain three segments of angle value, and restore the sampling direction code to the vehicle body sampling vector. S22. For each angle value, write the cosine of the angle value into the two coaxial holding positions of the rotation table, write the sine and negative sine of the angle value into the two non-axial reversing positions of the rotation table, write 1 into the rotation axis holding position of the rotation table, write zero into the remaining positions of the rotation table, and multiply the three rotation tables sequentially according to the three angle codes to generate a rotation matrix.

4. The fault diagnosis method for power inspection robots based on multimodal spatiotemporal alignment and three-layer fusion inference according to claim 3, characterized in that: S2 also includes: S23. Obtain the station observation vector by rotating the vehicle body sampling vector to the left using the rotation matrix, and read the component test end coordinates registered by the electrical performance testing device using the electrical component index of the suspicious word. S24. Subtract the current position of the inspection robot from the coordinates of the component test end to obtain the component direction vector. Subtract the component direction vector from the station observation vector to obtain the source verification difference vector. Write the source verification action word according to the axis sign of the source verification difference vector.

5. The fault diagnosis method for power inspection robots based on multimodal spatiotemporal alignment and three-layer fusion inference according to claim 4, characterized in that: S3 includes: S31. Based on the source action word, read the source difference vector and source action number. The edge computing unit recursively generates candidate actions according to the robot control resolution. For each candidate action generated, the source difference vector is obtained by subtracting the source difference vector from the sampled vector after the candidate action. The residual vector is then multiplied by itself and written into the action value. S32. For candidate actions corresponding to action values, the edge computing unit calculates the sign of the difference between the vehicle body position after the action and the boundary of the inspection channel, and calculates the sign of the difference between the test contact and the access point of the electrical performance test device after the action. If any difference sign exceeds the boundary, the candidate action is deleted, and the candidate actions are retained and the first item is written as the execution action word in ascending order of action value.

6. The fault diagnosis method for power inspection robots based on multimodal spatiotemporal alignment and three-layer fusion inference according to claim 5, characterized in that: S3 also includes: S33. Drive the inspection robot to change the sampling posture according to the execution action word. Recalculate the sampling vector with real-time Euler angles in each control cycle, and obtain the cycle residual vector by subtracting the source difference vector from the recalculated sampling vector. Write the sign of the difference between the sum of self-multiplication of the current cycle residual vector and the sum of self-multiplication of the previous cycle residual vector when the value changes from negative to zero. S34. At the action completion position, the edge computing unit triggers the electrical performance testing device to obtain the component electrical response with the electrical component index of the suspicious word, and triggers the electrical fault detection device to obtain the cable echo with the cable segment index of the suspicious word. Then, the component electrical response is connected to the cable echo and written as the re-collection word according to the source verification action number.

7. The fault diagnosis method for power inspection robots based on multimodal spatiotemporal alignment and three-layer fusion inference according to claim 6, characterized in that: S4 includes: S41. Read the suspicious word and the resampled word according to the sampling order of the source sequence. The neural encoder generates an input difference by subtracting the electrical measurement value of the previous word from the electrical measurement value of the current word. Multiply the input difference by the encoding weight, add the previous response code multiplied by the recursive weight, and write the summation value as the response code according to the encoding bit width. S42. Establish grid points by using the response code of the suspicious word as the row code and the response code of the repeated word as the column code. First, calculate the absolute value of the difference between the row code and the column code. Then, combine the sum of the bitwise XOR values ​​of the suspicious word object index and the repeated word object index into the absolute value of the difference to generate the grid point value.

8. The fault diagnosis method for power inspection robots based on multimodal spatiotemporal alignment and three-layer fusion inference according to claim 7, characterized in that: S4 also includes: S43. Recursively deduce the path table from the start position of the grid point to the end position of the grid point. For each grid point, read the cumulative value of the three predecessor grid points, take the first item in ascending order of value, add it to the current grid point value, and write it into the current cumulative value. Write the predecessor direction corresponding to the first item into the backtracking position. S44. After the path table is written to the end of the grid point, the grid point is read in reverse order from the end of the grid point according to the backtracking bit. When a grid point with a non-zero XOR sum of the object index is read, it is skipped. When a grid point with a zero XOR sum of the object index is read, it is written to the chain bit. The verification source alignment chain is generated in reverse order of the chain bit.

9. The fault diagnosis method for power inspection robots based on multimodal spatiotemporal alignment and three-layer fusion inference according to claim 8, characterized in that: S5 includes: S51. Read the suspicious word and the repeated word one by one along the verification source alignment chain. Subtract the electrical measurement response value of the suspicious word from the electrical measurement response value of the repeated word, and generate the response difference direction according to the sign of the difference. S52. Based on the source action word, read the source direction, multiply the response difference direction with the source direction bit by bit, delete the current chain bit when the product is negative, and write the current chain bit into the candidate chain when the product is non-negative.

10. The fault diagnosis method for power inspection robots based on multimodal spatiotemporal alignment and three-layer fusion inference according to claim 9, characterized in that: S5 also includes: S53. Read the electrical performance test value and fault detection value under the same object index for the candidate chain, multiply the sign of the difference between the electrical performance test value and the sign of the difference between the fault detection value, delete the current chain position when the product is negative, and write the current chain position into the verification chain when the product is non-negative. S54. Read the connection path corresponding to the object index in the verification chain along the electrical connection table, XOR the endpoint index of the connection path with the fault semantic index bit by bit, output the corresponding fault code when the XOR sum is zero, delete the current chain bit when the XOR sum is non-zero, and generate the inspection robot fault diagnosis result according to the retained fault code.

Images