Elevator fault classification method and system

Abstract

The invention relates to the technical field of fault classification, in particular to an elevator fault classification method and system.The elevator fault classification method comprises the steps that multi-source heterogeneous data, including mechanical vibration signals, driving motor three-phase current waveform signals, car video behavior data and control cabinet instruction streams, of elevator operation are collected in real time; performing space-time alignment processing on the multi-source heterogeneous data to generate a standardized time sequence data matrix; extracting multi-dimensional fault features from the standardized time sequence data matrix; constructing a dynamic knowledge graph based on a historical fault case library, wherein the knowledge graph comprises a three-layer topological structure of a fault feature node, a fault type node and a maintenance scheme node; and performing node matching on the extracted multi-dimensional fault features and the dynamic knowledge graph, and outputting a final fault classification result and confidence by adopting an improved residual neural network model. According to the method, the accuracy, real-time responsiveness and interpretability of fault diagnosis are remarkably improved, and the method has good engineering practical value and application and popularization prospects.

Description

Technical Field

[0001] The present invention relates to the technical field of fault classification, and in particular to an elevator fault classification method and system. Background Art

[0002] With the high-density development of urban buildings and the widespread deployment of smart buildings, the safety and stability of elevator systems, as an important part of vertical transportation, are increasingly valued. The operation and maintenance of existing elevators rely on periodic inspections and passive response mechanisms based on abnormal logs, which makes it difficult to detect potential fault risks in a timely manner. In recent years, although some systems have introduced sensors to collect vibration, current and other signals for preliminary diagnosis, most methods only rely on single modal data or static rule bases, which makes it difficult to achieve dynamic identification of complex fault situations. In particular, they are insufficient in the face of difficulties in fusing multi-source heterogeneous data, implicit expression of abnormal behavior, and co-occurrence of multiple types of faults. There are limitations in both accuracy and practicality.

[0003] In this context, there is an urgent need for an intelligent fault diagnosis solution that can integrate multimodal operation data, has spatiotemporal modeling capabilities and graph knowledge reasoning capabilities, so as to realize an end-to-end fault classification process from data perception, feature extraction to decision reasoning; combining knowledge graph technology and deep learning structure, it can embed historical case experience into the graph network in a structured manner, and realize semantic reasoning of the current operating status through the matching relationship between feature nodes and graph nodes; and then by introducing the residual neural network model, the expression ability of complex fault features and the classification confidence control ability are enhanced, thereby significantly improving the accuracy, real-time and interpretability of elevator fault detection. Summary of the invention

[0004] Based on the above purpose, the present invention provides an elevator fault classification method.

[0005] A method for classifying elevator faults comprises the following steps: S1: Real-time collection of multi-source heterogeneous data of elevator operation, including mechanical vibration signals, three-phase current waveform signals of drive motors, car video behavior data, and control cabinet command streams; S2: Performing spatiotemporal alignment processing on the multi-source heterogeneous data to generate a standardized time series data matrix; S3: Extract multi-dimensional fault features from the standardized time series data matrix, including the joint time-frequency domain features of the vibration signal, the harmonic distortion features of the current waveform, and the behavioral anomaly features of the video data; S4: construct a dynamic knowledge graph based on the historical fault case library, wherein the knowledge graph includes a three-layer topological structure of fault feature nodes, fault type nodes, and maintenance solution nodes; S5: Match the extracted multi-dimensional fault features with the dynamic knowledge graph nodes, dynamically adjust the graph node weights according to the matching degree, and use the improved residual neural network model to output the final fault classification results and confidence based on the optimized knowledge graph node weights.

[0006] Further, the S1 includes: S11, using a triaxial acceleration sensor installed on the surface of the elevator traction machine housing to collect mechanical vibration signals at a sampling rate of 1kHz; S12, collecting three-phase current waveform signals of the driving motor through a current transformer; S13, collecting passenger video behavior data in real time through a video collection device installed on the top of the car; S14, through accessing the CAN bus interface of the elevator control cabinet, collects the control cabinet command stream in real time.

[0007] Further, the S2 includes: S21, perform timestamp standardization on multi-source heterogeneous data; S22, constructing a global unified timeline based on the full set of timestamps in all data sources; S23, normalizing the alignment results of each data source; S24, concatenate all normalized data sources into a standardized time series data matrix.

[0008] Further, the S3 includes: S31, for vibration signals, short-time Fourier transform is used to extract its joint features in time and frequency domains; S32, extracting harmonic distortion characteristics of the current waveform using harmonic distortion rate for the current waveform signal; S33, for video data, using the frame difference method to extract abnormal behavior features and calculate the behavior motion energy; S34, the time-frequency domain joint features of the vibration signal, the harmonic distortion features of the current waveform, and the behavioral abnormality features of the video data are spliced ​​to form a multi-dimensional fault feature.

[0009] Furthermore, the time-frequency domain joint features of the vibration signal extracted in S31 include: S311, mapping the normalized mechanical vibration signal to the time-frequency domain using short-time Fourier transform to obtain an energy spectral density function; S312, extracting time series characteristic indicators from the energy spectrum, including main frequency distribution, instantaneous frequency, spectrum energy concentration, and frame energy; S313, statistically summarize the time series feature indicators based on the sliding window length to construct a joint feature vector in the time and frequency domains.

[0010] Further, the S4 includes: S41, extracting the fault feature vector, the corresponding fault type label and the maintenance plan information of each case based on the historical fault case library, and constructing a historical fault case set; S42, clustering the feature vectors in the historical data to generate fault feature nodes, combining the fault type labels and the maintenance plan records, respectively constructing fault type nodes and maintenance plan nodes, forming a three-layer topology structure including three types of nodes; S43, based on the similarity relationship between the fault feature nodes and the fault type nodes and the co-occurrence relationship between the fault type nodes and the maintenance solution nodes, directed edges are established in the graph and weights are assigned to construct a dynamic knowledge graph.

[0011] Further, the S5 includes: S51, node matching and weight adjustment: matching the extracted multi-dimensional fault features with the fault feature nodes in the knowledge graph, determining the relevance of the fault type nodes according to the matching results, and dynamically adjusting the weight of each fault type node in the knowledge graph; S52, fault classification reasoning output: Based on the adjusted knowledge graph node weights, the current fault features are fused with the graph weights and input into the improved residual neural network model to output the corresponding fault classification results and their confidence levels.

[0012] Further, the S51 includes: S511, using an exponential distance function to calculate a similarity score between the current feature vector and each fault feature node in the graph; S512, using the similarity score as the activation coefficient of the fault feature node, and passing the activation value to the fault type node set through the graph propagation mechanism to obtain the updated weight of each fault type node.

[0013] Further, the S52 includes: S521, concatenating the current fault feature vector with the adjusted fault type node weight vector to form a joint input vector; S522, using the joint input vector as an input of a residual neural network, where the network is composed of a plurality of stacked residual modules; The last layer output vector The data is input into the Softmax classifier, and the classification probability of each node in the fault type node set is calculated to obtain the final fault classification result and confidence.

[0014] An elevator fault classification system, used to implement the above-mentioned elevator fault classification method, includes the following modules: Data acquisition module: real-time acquisition of multi-source heterogeneous data of elevator operation, including mechanical vibration signals, three-phase current waveforms of drive motors, car video behavior data, and control cabinet command streams; Spatiotemporal alignment module: performs spatiotemporal alignment processing on the multi-source heterogeneous data to generate a standardized time series data matrix; Feature extraction module: extracts multi-dimensional fault features from the standardized time series data matrix, wherein the fault features include the joint time-frequency domain features of the vibration signal, the harmonic distortion features of the current waveform, and the abnormal behavior recognition features of the video data; Knowledge graph construction module: constructs a dynamic knowledge graph based on the historical fault case library, and the knowledge graph includes a three-layer topological structure of fault feature nodes, fault type nodes, and maintenance solution nodes; Graph matching and reasoning module: Perform node matching on the currently extracted multi-dimensional fault features and the dynamic knowledge graph, dynamically adjust the graph node weights according to the matching degree, and based on the optimized graph node weights, use the improved residual neural network model to output the fault classification results and their confidence.

[0015] Beneficial effects of the present invention: An elevator fault classification method provided by the present invention can integrate multi-source heterogeneous operation information such as mechanical vibration signals, current waveforms, video behavior data and control instructions, construct a standardized time series data matrix through a unified time-space alignment mechanism, and extract multi-dimensional features, including time-frequency joint features, harmonic distortion indicators and video behavior patterns. It overcomes the problems of traditional methods that rely on a single sensor signal and have weak recognition capabilities for complex abnormal behaviors, and effectively improves the comprehensiveness of feature expression and the sensitivity of fault characterization.

[0016] The three-layer dynamic knowledge graph constructed by the present invention based on historical cases can realize the structured modeling of the semantic relationship of elevator faults. Through the matching mechanism of fault feature nodes and graph nodes, combined with graph propagation and dynamic adjustment of node weights, the fusion reasoning of prior knowledge and current state can be realized. On this basis, the residual neural network structure is introduced to enhance the deep feature learning ability and improve the model stability. It can output classification results with confidence control, significantly improve the accuracy, real-time responsiveness and interpretability of fault diagnosis, and has good engineering practical value and application promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only for the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1is a flow chart of a method according to an embodiment of the present invention; Figure 2 4 is a system module diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments.

[0020] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0021] like Figure 1 As shown, a method for classifying elevator faults comprises the following steps: S1: Real-time collection of multi-source heterogeneous data of elevator operation, including mechanical vibration signals, three-phase current waveform signals of drive motors, car video behavior data, and control cabinet command streams; S2: Perform spatiotemporal alignment on multi-source heterogeneous data to generate a standardized time series data matrix; S3: Extract multi-dimensional fault features from the standardized time series data matrix, including the joint time-frequency domain features of the vibration signal, the harmonic distortion features of the current waveform, and the behavioral anomaly features of the video data; S4: Build a dynamic knowledge graph based on the historical fault case library. The knowledge graph includes a three-layer topology structure of fault feature nodes, fault type nodes, and maintenance solution nodes. S5: Match the extracted multi-dimensional fault features with the dynamic knowledge graph nodes, dynamically adjust the graph node weights according to the matching degree, and use the improved residual neural network model to output the final fault classification results and confidence based on the optimized knowledge graph node weights.

[0022] S1 includes: S11, a three-axis acceleration sensor is installed on the surface of the elevator traction machine housing to collect mechanical vibration signals at a sampling rate of 1kHz. The mechanical vibration signals are The vector at is represented as: ; in, Respectively indicate along The instantaneous acceleration component of the axis; S12, collect the three-phase current waveform signal of the driving motor through the current transformer. The three-phase current signal is at time The definition of is expressed as: ; in, Respectively represent motor , , The instantaneous current values ​​of the three phases are extracted based on the fast Fourier transform (FFT) to calculate the harmonic distortion (THD), which is expressed as: ; in, is the fundamental amplitude, For the The amplitude of the subharmonics, is the upper limit order of harmonic analysis; S13, collecting passenger video behavior data in real time through a video collection device installed on the top of the car; S14, through the CAN bus interface of the elevator control cabinet, collects the control cabinet command stream in real time, which is expressed as: ; in, Indicates Control commands, including start and stop instructions, elevator floor target, door control status, the sampling period is less than 100ms.

[0023] S2 includes: S21, standardize timestamps for multi-source heterogeneous data, sensors The collected original data sequence is expressed as: ; in, It is a sensor The original data sequence, Indicates The timestamp of each sampling point, represents the data vector corresponding to that moment, For sensor The number of sampling points; S22, construct a global unified timeline based on the full set of timestamps in all data sources, expressed as: ; in, To unify the aligned global timeline, is the target uniform time step number, and the time interval is , mapping all data sources to a unified time axis through linear interpolation, sensors At the moment The alignment value is expressed as: ; in, , indicating that the encirclement is found Linear interpolation is performed on the two original time points. is the interpolated sensor At the moment The alignment data value of S23, normalize the alignment results of each data source using the z-score normalization method, expressed as: ; in, is the aligned and normalized sensor At the moment The data value of Sensors The mean and standard deviation of all sampling points on the aligned time axis; S24, all normalized data sources are spliced ​​into a standardized time series data matrix, expressed as: ; in, Represents the total number of data sources, the final matrix , It is the total dimension after all data sources are concatenated.

[0024] S3 includes: S31, for vibration signals, short-time Fourier transform (STFT) is used to extract its joint features in time and frequency domains, which can be expressed as: ; in, is the vibration signal under the unified time axis, is the time variable used for short-time Fourier transform (STFT) integration or summation, For time The sliding window function (Hamming window) centered at is the frequency variable, Indicates time frequency The local spectrum response under Construct a time-frequency energy spectrum and extract its main frequency distribution, instantaneous frequency, and energy center features as a joint feature vector in the time-frequency domain ; S32, for the current waveform signal, the harmonic distortion rate (THD) is used to extract the harmonic distortion characteristics of the current waveform, which is expressed as: ; in, is the three-phase current signal, , Current signal No. The amplitude of the subharmonic is obtained by FFT transformation. is the fundamental amplitude, is the highest order of harmonic analysis, and the current eigenvector is ; S33, for video data, the frame difference method is used to extract abnormal behavior features. The grayscale difference between two consecutive frames of images is expressed as: ; in, are the image pixel coordinates, It is Frame at pixel location The gray value of is the brightness change between the current frame and the previous frame at the pixel position. and It is two consecutive frames of images; Calculate the energy of the behavior , expressed as: ; Constructing motion energy sequence using sliding window method , extract energy mean, peak value and mutation point as video behavior feature vector ; S34, the time-frequency domain joint features of the vibration signal, the harmonic distortion features of the current waveform, and the behavioral abnormality features of the video data are spliced ​​to form a multi-dimensional fault feature, which is expressed as: .

[0025] The time-frequency domain joint features of the vibration signal extracted in S31 include: S311, use short-time Fourier transform (STFT) to map the normalized mechanical vibration signal to the time-frequency domain and obtain the energy spectral density function , expressed as: ; S312, extracting time series characteristic indicators from the energy spectrum, including main frequency distribution, instantaneous frequency, spectrum energy concentration and frame energy, expressed as: 1) Main frequency distribution : ; 2) Instantaneous frequency (spectral center): ; 3) Spectral energy concentration (spectral abstract) : ； ; 4) Frame Energy : ; S313, with sliding window length The time series feature indicators are statistically summarized and the joint feature vector in the time and frequency domain is constructed, which is expressed as: ; in, Characteristics Mean and standard deviation within the window.

[0026] S4 includes: S41, based on the historical fault case library, extract the fault feature vector, the corresponding fault type label and the maintenance plan information of each case, and construct a historical fault case set, which is expressed as: ; in, Indicates 'Historical failure cases, including multi-dimensional failure feature vectors , corresponding to the fault type label , Maintenance plan description ; S42, clustering the feature vectors in the historical data to generate fault feature nodes, combining the fault type labels and maintenance plan records, respectively constructing fault type nodes and maintenance plan nodes, forming a three-layer topology structure including three types of nodes, specifically including: (1) Fault feature node set : ; in, Represents the subspace center of different types of fault characteristics (obtained by clustering), Indicates The representation vector of feature nodes; (2) Fault type node set : ; in, represents the fault classification labels known in history, Indicates The representation vector of the fault type nodes; (3) Maintenance plan node set : ; in, Represents the standardized maintenance plan descriptions associated with each type of fault, Indicates The representation vector of the maintenance solution node; S43, based on the similarity relationship between the fault feature node and the fault type node and the co-occurrence relationship between the fault type node and the maintenance solution node, a directed edge is established in the graph and a weight is assigned to construct a dynamic knowledge graph, which is expressed as: ; Among them, the fault type node set is , the edge set is , the steps of establishing edge sets specifically include: (1) For fault feature nodes Node with fault type , calculate the matching edge weight according to the Euclidean distance, expressed as: ; in, Fault Type The corresponding feature mean vector, is the characteristic distance scale parameter; (2) For fault type nodes Maintenance plan node , the edge weight is obtained by normalizing the historical co-occurrence times, expressed as: ; in, Fault Type Maintenance plan node Co-occurrence frequency in the case base.

[0027] S5 includes: S51, node matching and weight adjustment: matching the extracted multi-dimensional fault features with the fault feature nodes in the knowledge graph, determining the relevance of the fault type nodes according to the matching results, and dynamically adjusting the weight of each fault type node in the knowledge graph; S52, fault classification reasoning output: Based on the adjusted knowledge graph node weights, the current fault features are fused with the graph weights and input into the improved residual neural network model to output the corresponding fault classification results and their confidence levels.

[0028] S51 includes: S511, using exponential distance function to calculate the current feature vector With each fault feature node in the graph Similarity score , expressed as: ; in, Adjust parameters for matching scale; S512, the similarity score is used as the activation coefficient of the fault feature node, and the activation value is transmitted to the fault type node set through the graph propagation mechanism. , and obtain the updated weight of each fault type node, expressed as: ; in, Fault feature node Node with fault type The edge weights between Indicates the fault type node The dynamically adjusted weights.

[0029] S52 includes: S521, compare the current fault feature vector with the adjusted fault type node weight vector Concatenate to form a joint input vector , expressed as: ; S522, the joint input vector is used as the input of the residual neural network, the network is composed of multiple stacked residual modules, and the output definition of each residual module is expressed as: ; in, and Respectively The weight matrix and bias vector of the layer, is a nonlinear activation function; The last layer output vector Input to the Softmax classifier, and calculate the classification probability of each node in the fault type node set, expressed as: ; in, Fault type node The classification confidence of and are their classification weight and bias respectively; The final fault classification result and confidence are obtained, expressed as: ; .

[0030] like Figure 2 As shown, an elevator fault classification system is used to implement the above-mentioned elevator fault classification method, including the following modules: Data acquisition module: real-time acquisition of multi-source heterogeneous data of elevator operation, including mechanical vibration signals, three-phase current waveforms of drive motors, car video behavior data, and control cabinet command streams; Spatiotemporal alignment module: performs spatiotemporal alignment processing on multi-source heterogeneous data to generate a standardized time series data matrix; Feature extraction module: extracts multi-dimensional fault features from the standardized time series data matrix. The fault features include the joint time-frequency domain features of the vibration signal, the harmonic distortion features of the current waveform, and the abnormal behavior recognition features of the video data. Knowledge graph construction module: constructs a dynamic knowledge graph based on the historical fault case library. The knowledge graph includes a three-layer topology structure of fault feature nodes, fault type nodes, and maintenance solution nodes; Graph matching and reasoning module: Perform node matching on the currently extracted multi-dimensional fault features and the dynamic knowledge graph, dynamically adjust the graph node weights according to the matching degree, and based on the optimized graph node weights, use the improved residual neural network model to output the fault classification results and their confidence.

[0031] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of the present invention is limited to these examples. Under the concept of the present invention, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of the different aspects of the present invention as described above, which are not provided in detail for the sake of simplicity.

Claims

1. A method for classifying elevator faults, characterized in that: The following steps are involved: S1: Real-time collection of multi-source heterogeneous data of elevator operation, including mechanical vibration signals, three-phase current waveform signals of drive motors, car video behavior data, and control cabinet command streams; S2: Performing spatiotemporal alignment processing on the multi-source heterogeneous data to generate a standardized time series data matrix; S3: Extract multi-dimensional fault features from the standardized time series data matrix, including the joint time-frequency domain features of the vibration signal, the harmonic distortion features of the current waveform, and the behavioral anomaly features of the video data; S4: construct a dynamic knowledge graph based on the historical fault case library, wherein the knowledge graph includes a three-layer topological structure of fault feature nodes, fault type nodes, and maintenance solution nodes; S5: Match the extracted multi-dimensional fault features with the dynamic knowledge graph nodes, dynamically adjust the graph node weights according to the matching degree, and use the improved residual neural network model to output the final fault classification results and confidence based on the optimized knowledge graph node weights.

2. An elevator fault classification method according to claim 1, characterized in that: The S1 includes: S11, using a triaxial acceleration sensor installed on the surface of the elevator traction machine housing to collect mechanical vibration signals at a sampling rate of 1kHz; S12, collecting three-phase current waveform signals of the driving motor through a current transformer; S13, collecting passenger video behavior data in real time through a video collection device installed on the top of the car; S14, through accessing the CAN bus interface of the elevator control cabinet, collects the control cabinet command stream in real time.

3. An elevator fault classification method according to claim 2, characterized in that: The S2 includes: S21, perform timestamp standardization on multi-source heterogeneous data; S22, constructing a global unified timeline based on the full set of timestamps in all data sources; S23, normalizing the alignment results of each data source; S24, concatenate all normalized data sources into a standardized time series data matrix.

4. An elevator fault classification method according to claim 3, characterized in that: The S3 includes: S31, for vibration signals, short-time Fourier transform is used to extract its joint features in time and frequency domains; S32, extracting harmonic distortion characteristics of the current waveform using harmonic distortion rate for the current waveform signal; S33, for video data, using the frame difference method to extract abnormal behavior features and calculate the behavior motion energy; S34, the time-frequency domain joint features of the vibration signal, the harmonic distortion features of the current waveform, and the behavioral abnormality features of the video data are spliced ​​to form a multi-dimensional fault feature.

5. An elevator fault classification method according to claim 4, characterized in that: Extracting the time-frequency domain joint features of the vibration signal in S31 includes: S311, mapping the normalized mechanical vibration signal to the time-frequency domain using short-time Fourier transform to obtain an energy spectral density function; S312, extracting time series characteristic indicators from the energy spectrum, including main frequency distribution, instantaneous frequency, spectrum energy concentration, and frame energy; S313, statistically summarize the time series feature indicators based on the sliding window length to construct a joint feature vector in the time and frequency domains.

6. An elevator fault classification method according to claim 5, characterized in that: The S4 includes: S41, extracting the fault feature vector, the corresponding fault type label and the maintenance plan information of each case based on the historical fault case library, and constructing a historical fault case set; S42, clustering the feature vectors in the historical data to generate fault feature nodes, combining the fault type labels and the maintenance plan records, respectively constructing fault type nodes and maintenance plan nodes, forming a three-layer topology structure including three types of nodes; S43, based on the similarity relationship between the fault feature nodes and the fault type nodes and the co-occurrence relationship between the fault type nodes and the maintenance solution nodes, directed edges are established in the graph and weights are assigned to construct a dynamic knowledge graph.

7. An elevator fault classification method according to claim 6, characterized in that: The S5 includes: S51, node matching and weight adjustment: matching the extracted multi-dimensional fault features with the fault feature nodes in the knowledge graph, determining the relevance of the fault type nodes according to the matching results, and dynamically adjusting the weight of each fault type node in the knowledge graph; S52, fault classification reasoning output: Based on the adjusted knowledge graph node weights, the current fault features are fused with the graph weights and input into the improved residual neural network model to output the corresponding fault classification results and their confidence levels.

8. An elevator fault classification method according to claim 7, characterized in that: The S51 includes: S511, using an exponential distance function to calculate a similarity score between the current feature vector and each fault feature node in the graph; S512, using the similarity score as the activation coefficient of the fault feature node, and passing the activation value to the fault type node set through the graph propagation mechanism to obtain the updated weight of each fault type node.

9. An elevator fault classification method according to claim 7, characterized in that: The S52 includes: S521, concatenating the current fault feature vector with the adjusted fault type node weight vector to form a joint input vector; S522, using the joint input vector as an input of a residual neural network, where the network is composed of a plurality of stacked residual modules; The last layer output vector The data is input into the Softmax classifier, and the classification probability of each node in the fault type node set is calculated to obtain the final fault classification result and confidence.

10. An elevator fault classification system, used to implement an elevator fault classification method according to any one of claims 1 to 9, characterized in that: Includes the following modules: Data acquisition module: real-time acquisition of multi-source heterogeneous data of elevator operation, including mechanical vibration signals, three-phase current waveforms of drive motors, car video behavior data, and control cabinet command streams; Spatiotemporal alignment module: performs spatiotemporal alignment processing on the multi-source heterogeneous data to generate a standardized time series data matrix; Feature extraction module: extracts multi-dimensional fault features from the standardized time series data matrix, wherein the fault features include the joint time-frequency domain features of the vibration signal, the harmonic distortion features of the current waveform, and the abnormal behavior recognition features of the video data; Knowledge graph construction module: constructs a dynamic knowledge graph based on the historical fault case library, and the knowledge graph includes a three-layer topological structure of fault feature nodes, fault type nodes, and maintenance solution nodes; Graph matching and reasoning module: Perform node matching on the currently extracted multi-dimensional fault features and the dynamic knowledge graph, dynamically adjust the graph node weights according to the matching degree, and based on the optimized graph node weights, use the improved residual neural network model to output the fault classification results and their confidence.

Images