Methods, devices, equipment, and storage media for diagnosing anomalies in substation power equipment
By extracting datasets based on the time of anomalies in substation power equipment anomaly diagnosis and using a diagnostic model for cross-modal matching, the problem of time synchronization between acoustic print data and infrared data is solved, thus improving the accuracy and interpretability of anomaly diagnosis.
Patent Information
- Application Number
- CN202611602363.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-16
- Publication Date
- 2026-11-17
AI Technical Summary
In existing technologies, time synchronization between voiceprint data and infrared data is difficult, which affects the accuracy of abnormal diagnosis of substation power equipment.
By using the abnormal time corresponding to the abnormal features as a benchmark, the running dataset within a preset time period is extracted, and the acoustic abnormal events and thermal abnormal events are matched using a pre-built diagnostic model. A cross-modal matching algorithm is used to obtain the quality of matched event pairs and unmatched events, and the data is fused to obtain the abnormal evolution state and diagnostic results.
This effectively avoids the difficulty of point-to-point alignment between voiceprint data and infrared data, improves the accuracy and interpretability of abnormal diagnosis of substation power equipment, and ensures that single-modal abnormal evidence is not discarded.
Smart Images

Figure CN123122006A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of power equipment condition monitoring technology, and in particular to a method, apparatus, equipment and storage medium for abnormal diagnosis of power equipment in substations. Background Technology
[0002] With the development of intelligent condition monitoring technology for substations, online monitoring systems for power equipment typically need to collect multi-source data such as acoustic signatures, infrared sensors, and operating conditions, and then fuse and analyze these data to achieve equipment anomaly diagnosis. Therefore, a power equipment anomaly diagnosis technology based on cross-modal fusion of acoustic signatures and infrared sensors has emerged. This technology extracts acoustic signature anomaly features and infrared thermal anomaly features separately, and then correlates these two types of features across modalities to determine the abnormal state of the equipment.
[0003] Existing technologies mainly employ a combination of time synchronization and independent diagnostics. This involves aligning voiceprint data and infrared data along a unified timeline, then independently detecting anomalies in the voiceprint data and infrared data, and finally matching the detection results by timestamp to determine if the device is malfunctioning.
[0004] However, this method requires strict time synchronization between voiceprint data and infrared data. Voiceprint is a high-frequency continuous signal, while infrared is a low-frame-rate image. The sampling rates of the two are very different, making time alignment very difficult and prone to data misalignment, which in turn seriously affects the accuracy of equipment anomaly diagnosis. Summary of the Invention
[0005] Therefore, it is necessary to provide a method, device, equipment, and storage medium for diagnosing abnormalities in substation power equipment to address the aforementioned technical problems.
[0006] Firstly, this application provides a method for diagnosing abnormalities in substation power equipment, including:
[0007] The system acquires operational data of the target power equipment. When abnormal features are detected in the operational data, the system extracts the operational dataset within a preset time period based on the abnormal time corresponding to the abnormal features. The operational data includes acoustic data and infrared data.
[0008] The running dataset is input into a pre-built diagnostic model, which extracts acoustic and thermal anomalies. Based on a preset matching algorithm, the acoustic and thermal anomalies are matched to obtain matched event pairs and unmatched event quality. The matched event pairs and unmatched event quality are then fused to obtain the abnormal evolution state and diagnostic results of the target power equipment.
[0009] In one embodiment, the diagnostic model includes a normal response model, a prediction model, a matching model, and a replacement model. The running dataset is input into the pre-built diagnostic model, which extracts acoustic and thermal anomalies. Based on a preset matching algorithm, the acoustic and thermal anomalies are matched to obtain matched event pairs and unmatched event qualities. The matched event pairs and unmatched event qualities are then fused to obtain the abnormal evolution state and diagnostic results of the target power equipment, including:
[0010] The acoustic and infrared data from the running dataset are input into the normal response model, which generates acoustic and infrared residuals, and then extracts acoustic and thermal anomaly events from the acoustic and infrared residuals.
[0011] The acoustic anomaly events are input into the prediction model to obtain the prediction results. The total mass of the acoustic events and the total mass of the thermal events are obtained by combining the prediction results, the thermal anomaly events and the pre-determined hierarchical confidence.
[0012] The total mass of acoustic events and the total mass of thermal events are input into the matching model, which performs cross-modal matching according to the matching algorithm to obtain the matched event pairs, the mass of unmatched acoustic events, the mass of unmatched thermal events, and the cross-modal association confidence.
[0013] The matched event pairs, the quality of the unmatched acoustic events, the quality of the unmatched thermal events, and the cross-modal correlation confidence are input into the replacement model. The replacement model replaces the acoustic residuals with the normal acoustic response under the current operating conditions and the infrared residuals with the normal infrared response under the current operating conditions. The changes in the diagnostic conclusions before and after the replacement are calculated to obtain the voiceprint evidence contribution and the infrared evidence contribution. The matched event pairs, the quality of the unmatched acoustic events, the quality of the unmatched thermal events, the cross-modal correlation confidence, the hierarchical confidence, the voiceprint evidence contribution, and the infrared evidence contribution are fused to obtain the abnormal evolution state and the diagnostic results.
[0014] In one embodiment, cross-modal matching is performed according to a matching algorithm to obtain matched event pairs, the quality of unmatched acoustic events, the quality of unmatched thermal events, and the cross-modal association confidence level, including:
[0015] The matching cost is determined based on the time delay distribution, the characteristic distance between acoustic and thermal anomalies, the topological relationship between the corresponding equipment components of acoustic and thermal anomalies, and the local reliability of acoustic and thermal anomalies.
[0016] Cross-modal matching is performed on the total mass of acoustic events and the total mass of thermal events based on the matching algorithm and matching cost, to obtain the matched event pairs, the mass of unmatched acoustic events, the mass of unmatched thermal events, and the cross-modal association confidence.
[0017] In one embodiment, cross-modal matching is performed on the total acoustic event quality and the total thermal event quality based on a matching algorithm and a matching cost to obtain matched event pairs, unmatched acoustic event quality, unmatched thermal event quality, and cross-modal association confidence, including:
[0018] Based on the matching algorithm, with the matching cost as the objective, the total mass of acoustic events and the total mass of thermal events are matched to obtain the matching matrix between acoustic anomaly events and thermal anomaly events. Each element in the matching matrix represents the matching mass between each acoustic anomaly event and thermal anomaly event. Acoustic anomaly events and thermal anomaly events that have not been assigned matching mass remain unmatched.
[0019] Based on the matching matrix, the acoustic event quality corresponding to the unmatched acoustic anomaly event is determined as the unmatched acoustic event quality, the thermal event quality corresponding to the unmatched thermal anomaly event is determined as the unmatched thermal event quality, and the acoustic anomaly event and thermal anomaly event with matching quality higher than a preset threshold are determined as matched event pairs.
[0020] The reliability of cross-modal association is determined based on the matching quality and corresponding matching cost of each event pair among all matched event pairs.
[0021] In one embodiment, the layered reliability includes acoustic modal reliability, infrared modal reliability, acoustic event local reliability, and thermal event local reliability; the method for determining the layered reliability includes:
[0022] The credibility of the voiceprint modality is determined based on the quality indicators of the voiceprint data.
[0023] The reliability of infrared modes is determined based on the quality indicators of infrared data.
[0024] Determine the local reliability of acoustic events based on the event quality index of acoustic anomalies;
[0025] The local reliability of a thermal event is determined based on the event quality index of the thermal anomaly.
[0026] In one embodiment, acoustic anomalies are input into a prediction model to obtain prediction results. The total acoustic event mass and total thermal event mass are then obtained by combining the prediction results, thermal anomalies, and pre-determined hierarchical confidence levels, including:
[0027] The acoustic anomaly event is input into the prediction model for prediction, and the time delay distribution and waiting probability corresponding to the acoustic anomaly event are obtained. The waiting probability is used to characterize the probability that the corresponding thermal anomaly event will occur after the current time.
[0028] The total mass of the acoustic event is determined based on the time delay distribution, waiting probability, and hierarchical credibility.
[0029] The total mass of a thermal event is determined based on the infrared modal reliability of the layered reliability and the local reliability of the thermal event in the thermal anomaly.
[0030] In one embodiment, the matched event pairs, the quality of unmatched acoustic events, the quality of unmatched thermal events, the cross-modal association confidence, the hierarchical confidence, the contribution of acoustic evidence, and the contribution of infrared evidence are fused to obtain the abnormal evolution state and diagnostic results, including:
[0031] Based on the matched event pairs, the matched acoustic anomaly events and thermal anomaly events are weighted and aggregated to obtain the features of the matched event pairs;
[0032] Unmatched acoustic and thermal anomaly events are weighted and aggregated based on their quality to obtain unmatched single-modal features.
[0033] The matching event pair features, unmatched single-modal features, hierarchical credibility, cross-modal association credibility, voiceprint evidence contribution, and infrared evidence contribution are concatenated to obtain the fused features;
[0034] The fusion features are processed for state discrimination to obtain abnormal evolution states and diagnostic results.
[0035] Secondly, this application also provides an anomaly diagnosis device for substation power equipment, comprising:
[0036] The acquisition module is used to acquire the operating data of the target power equipment. When abnormal features are detected in the operating data, the operating data set within a preset time period is extracted based on the abnormal time corresponding to the abnormal features. The operating data includes voiceprint data and infrared data.
[0037] The input module is used to input the running dataset into the pre-built diagnostic model. The diagnostic model extracts acoustic and thermal anomalies, matches the acoustic and thermal anomalies based on a preset matching algorithm, obtains the quality of matched event pairs and unmatched events, and performs fusion processing based on the quality of matched event pairs and unmatched events to obtain the abnormal evolution state and diagnostic results of the target power equipment.
[0038] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:
[0039] The system acquires operational data of the target power equipment. When abnormal features are detected in the operational data, the system extracts the operational dataset within a preset time period based on the abnormal time corresponding to the abnormal features. The operational data includes acoustic data and infrared data.
[0040] The running dataset is input into a pre-built diagnostic model, which extracts acoustic and thermal anomalies. Based on a preset matching algorithm, the acoustic and thermal anomalies are matched to obtain matched event pairs and unmatched event quality. The matched event pairs and unmatched event quality are then fused to obtain the abnormal evolution state and diagnostic results of the target power equipment.
[0041] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:
[0042] The system acquires operational data of the target power equipment. When abnormal features are detected in the operational data, the system extracts the operational dataset within a preset time period based on the abnormal time corresponding to the abnormal features. The operational data includes acoustic data and infrared data.
[0043] The running dataset is input into a pre-built diagnostic model, which extracts acoustic and thermal anomalies. Based on a preset matching algorithm, the acoustic and thermal anomalies are matched to obtain matched event pairs and unmatched event quality. The matched event pairs and unmatched event quality are then fused to obtain the abnormal evolution state and diagnostic results of the target power equipment.
[0044] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:
[0045] The system acquires operational data of the target power equipment. When abnormal features are detected in the operational data, the system extracts the operational dataset within a preset time period based on the abnormal time corresponding to the abnormal features. The operational data includes acoustic data and infrared data.
[0046] The running dataset is input into a pre-built diagnostic model, which extracts acoustic and thermal anomalies. Based on a preset matching algorithm, the acoustic and thermal anomalies are matched to obtain matched event pairs and unmatched event quality. The matched event pairs and unmatched event quality are then fused to obtain the abnormal evolution state and diagnostic results of the target power equipment.
[0047] The aforementioned method, device, equipment, and storage medium for diagnosing anomalies in substation power equipment acquires the operating data of the target power equipment. When anomalies are detected in the operating data, the operating dataset within a preset time period is extracted based on the anomaly time corresponding to the anomaly. The operating data includes acoustic fingerprint data and infrared data. The operating dataset is input into a pre-built diagnostic model, which extracts acoustic and thermal anomalies. Based on a preset matching algorithm, the acoustic and thermal anomalies are matched to obtain matched event pairs and unmatched event quality. The matched event pairs and unmatched event quality are then fused to obtain the anomaly evolution state and diagnostic results of the target power equipment. The above method extracts the operational dataset within a preset time period based on the abnormal time corresponding to the abnormal features, aligning the acoustic and infrared data around the abnormal events and avoiding the difficulty of point-by-point alignment. It matches acoustic and thermal abnormal events based on a preset matching algorithm to obtain matched event pairs and unmatched event quality, ensuring that single-modal abnormal evidence is not discarded. By fusing the matched event pairs and unmatched event quality, the accuracy and interpretability of substation power equipment anomaly diagnosis are improved. Attached Figure Description
[0048] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0049] Figure 1 This is a diagram illustrating the application environment of a substation power equipment anomaly diagnosis method in one embodiment.
[0050] Figure 2 This is one of the flowcharts illustrating an anomaly diagnosis method for substation power equipment in one embodiment;
[0051] Figure 3 This is a second flowchart illustrating a method for diagnosing abnormalities in substation power equipment in one embodiment.
[0052] Figure 4 This is the third flowchart of a method for diagnosing abnormalities in substation power equipment in one embodiment;
[0053] Figure 5 This is the fourth flowchart illustrating a method for diagnosing abnormalities in substation power equipment in one embodiment;
[0054] Figure 6This is the fifth flowchart illustrating a method for diagnosing abnormalities in substation power equipment in one embodiment;
[0055] Figure 7 This is a structural block diagram of an anomaly diagnosis device for substation power equipment in one embodiment. Detailed Implementation
[0056] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0057] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.
[0058] With the development of intelligent condition monitoring technology for substations, online monitoring systems for power equipment typically need to collect multi-source data such as acoustic signatures, infrared sensors, and operating conditions, and then fuse and analyze these data to achieve equipment anomaly diagnosis. Therefore, a power equipment anomaly diagnosis technology based on cross-modal fusion of acoustic signatures and infrared sensors has emerged. This technology extracts acoustic signature anomaly features and infrared thermal anomaly features separately, and then correlates these two types of features across modalities to determine the abnormal state of the equipment.
[0059] Existing technologies mainly employ a combination of time synchronization and independent diagnostics. This involves aligning voiceprint data and infrared data along a unified timeline, then independently detecting anomalies in the voiceprint data and infrared data, and finally matching the detection results by timestamp to determine if the device is malfunctioning.
[0060] However, this method requires strict time synchronization between voiceprint data and infrared data. Voiceprint is a high-frequency continuous signal, while infrared is a low-frame-rate image. The sampling rates of the two are very different, making time alignment very difficult and prone to data misalignment, which in turn seriously affects the accuracy of equipment anomaly diagnosis.
[0061] In view of the above-mentioned technical problems, this application provides a method for diagnosing abnormalities in substation power equipment. The following embodiments will specifically illustrate the method for diagnosing abnormalities in substation power equipment.
[0062] The abnormal diagnosis method for substation power equipment provided in this application embodiment can be applied to, for example... Figure 1The computer device shown includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When the computer program is executed by the processor, it implements a method for diagnosing abnormalities in substation power equipment. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.
[0063] Those skilled in the art will understand that Figure 1 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0064] In one exemplary embodiment, such as Figure 2 As shown, a method for diagnosing anomalies in substation power equipment is provided, which can be applied to... Figure 1 Taking computer devices as an example, the explanation includes:
[0065] S201, acquire the operating data of the target power equipment, and if abnormal features are detected in the operating data, extract the operating dataset within a preset time period based on the abnormal time corresponding to the abnormal features.
[0066] The operational data includes acoustic fingerprint data and infrared data. Acoustic fingerprint data is a high-frequency continuous signal sequence collected by an acoustic fingerprint sensor, used to reflect acoustic anomalies such as mechanical vibration, frictional impact, and partial discharge of the target power equipment; infrared data is a low-frame-rate temperature matrix or infrared image sequence collected by an infrared camera, used to reflect thermal anomalies such as poor contact heating, abnormal heat dissipation, and load-related temperature rise of the target power equipment.
[0067] In this embodiment, the computer device acquires operational data reported by the target power equipment in real time. The operational data includes acoustic fingerprint data X_a(t) and infrared data X_i(k), where t is the time index of the acoustic fingerprint data and k is the frame index of the infrared data. Subsequently, the computer device performs anomaly detection on the operational data. When an anomaly is detected, the computer extracts the operational dataset within a preset time period before and after the anomaly, using the anomaly time t_0 corresponding to that anomaly as a baseline. The anomaly can be an acoustic anomaly peak in the acoustic fingerprint data, a temperature rise inflection point in the infrared data, a load mutation event, or an alarm event. The operational dataset within the preset time period is a data segment extended forward and backward along the time axis from the anomaly time t_0, used to preserve complete contextual information before and after the occurrence of the anomaly.
[0068] Since voiceprint data is usually a high-frequency continuous sequence and infrared data is usually a low-frame-rate image or temperature matrix, it is very difficult to directly align the running dataset point by point. Therefore, the computer device constructs an asynchronous diagnostic sample set around the device events. Specifically, let the voiceprint acquisition interval be [t_0, t_0+T_a] and the infrared acquisition interval be [t_0-_pre, t_0+T_i+_post], then the diagnostic sample X can be represented by the relation (1), which is shown below:
[0069] X={X_a(t),X_i(k),X_c(t),e,G_d,[-_pre,_post]}(1);
[0070] Where T_a is the duration of the acoustic signature acquisition interval; T_i is the reference duration of the infrared acquisition interval; _pre and _post are the allowed forward and backward time delay ranges, respectively, that is, the duration of the infrared acquisition interval extending forward and backward relative to the abnormal moment, used to cover the possible physical time delay between the acoustic signature anomaly and the infrared thermal response; X_c(t) is the operating condition data; e is the equipment or component identifier; G_d is the equipment or component topology.
[0071] In some embodiments, time correction can be achieved using a uniform clock, device operation events, load surge events, alarm events, acoustic anomaly peaks, infrared temperature rise inflection points, or sliding window matching. When the precise time offset cannot be determined, the offset probability distribution is preserved, and a unique time offset is not mandatory.
[0072] In some embodiments, the operational data includes not only acoustic fingerprint data and infrared data, but also operating condition data. Equipment operating condition data may include load current, voltage, active / reactive power, operating status, ambient temperature, humidity, wind speed, and sunshine duration. Optionally, the operational data may also include equipment or component identification and equipment / component topology.
[0073] In some embodiments, each diagnostic sample X maintains both modal integrity information and local validity information. For example, when the infrared target region is partially occluded, only the local reliability of events in the relevant target region is reduced; when the entire infrared temperature matrix is incomplete, the global reliability of the infrared modes is further reduced.
[0074] S202, the running dataset is input into the pre-built diagnostic model, the diagnostic model extracts acoustic anomalies and thermal anomalies, the acoustic anomalies and thermal anomalies are matched based on the preset matching algorithm, the matched event pairs and unmatched event quality are obtained, and the matched event pairs and unmatched event quality are fused to obtain the abnormal evolution state and diagnostic results of the target power equipment.
[0075] Among them, acoustic anomalies are local anomalies extracted from voiceprint data that deviate from the normal acoustic response; thermal anomalies are local anomalies extracted from infrared data that deviate from the normal infrared response.
[0076] In this embodiment, after obtaining an asynchronous diagnostic sample set, the computer device inputs this asynchronous diagnostic sample set as a running dataset into a pre-built diagnostic model. The diagnostic model processes the running dataset to obtain the abnormal evolution state and diagnostic results of the target power equipment. Specifically, the diagnostic model includes a normal response model, a prediction model, a matching model, and a replacement model. The normal response model generates acoustic residuals and infrared residuals based on the acoustic and infrared data in the running dataset, and extracts acoustic and thermal anomalies from them. Subsequently, the prediction unit predicts the acoustic anomalies and determines the total acoustic and thermal event quality based on hierarchical confidence. Next, the matching model performs cross-modal matching on the total acoustic and thermal event quality according to a preset matching algorithm, obtaining matched event pairs, unmatched acoustic event quality, unmatched thermal event quality, and cross-modal association confidence. Finally, the replacement model determines the acoustic and infrared evidence contributions through normal prototype replacement based on the matched event pairs, unmatched acoustic and thermal event quality, and cross-modal association confidence, and fuses them to obtain the abnormal evolution state and diagnostic results.
[0077] In some embodiments, the abnormal evolution state includes at least one of the following: normal, acoustic precursor pending correlation, single-mode acoustic anomaly, single-mode thermal anomaly, thermal anomaly formation, acoustic-thermal synergistic anomaly, suspected composite / conflict anomaly, or insufficient data quality requiring review. The diagnostic results also include anomaly type tendencies and interpretable information. Anomaly type tendencies include at least one of the following: mechanical vibration, friction / impact, mechanical loosening, suspected partial discharge, poor contact / heating at connection points, abnormal heat dissipation / cooling, load-related temperature rise anomaly, or unknown anomaly.
[0078] In the above-mentioned method for diagnosing anomalies in substation power equipment, the operating data of the target power equipment is acquired. When anomalies are detected in the operating data, an operating dataset within a preset time period is extracted based on the anomaly time corresponding to the anomaly. The operating data includes acoustic fingerprint data and infrared data. The operating dataset is input into a pre-built diagnostic model, which extracts acoustic and thermal anomalies. Based on a preset matching algorithm, the acoustic and thermal anomalies are matched to obtain matched event pairs and unmatched event quality. The matched event pairs and unmatched event quality are then fused to obtain the anomaly evolution state and diagnostic result of the target power equipment. This method extracts the operating dataset within a preset time period based on the anomaly time corresponding to the anomaly, aligning the acoustic fingerprint and infrared data around the anomaly events, thus avoiding the difficulty of point-by-point alignment. The matching algorithm based on the preset matching algorithm ensures that single-modal anomaly evidence is not discarded. The fusion of matched event pairs and unmatched event quality improves the accuracy and interpretability of anomaly diagnosis in substation power equipment.
[0079] In an exemplary embodiment, the diagnostic model includes a normal response model, a prediction model, a matching model, and a replacement model. The above-mentioned S202 step of "inputting the running dataset into the pre-built diagnostic model, extracting acoustic and thermal anomaly events by the diagnostic model, matching the acoustic and thermal anomaly events based on a preset matching algorithm to obtain matched event pairs and unmatched event quality, and performing fusion processing based on the matched event pairs and unmatched event quality to obtain the abnormal evolution state and diagnostic results of the target power equipment" is as follows: Figure 3 As shown, it includes:
[0080] S301, input the acoustic fingerprint data and infrared data in the running dataset into the normal response model respectively, generate acoustic residuals and infrared residuals from the normal response model, and extract acoustic anomaly events and thermal anomaly events from the acoustic residuals and infrared residuals.
[0081] Among them, acoustic residuals are used to characterize anomalous components in acoustic data that deviate from the normal acoustic response, and infrared residuals are used to characterize anomalous components in infrared data that deviate from the normal infrared response.
[0082] In this embodiment, the computer device inputs the acoustic fingerprint data and infrared data from the running dataset into a pre-built normal response model. The normal response model includes an acoustic normal response model B_a and an infrared normal thermal field model B_i, which can be represented by relation (2), as shown below:
[0083] F_hat_a^0(t)=B_a(e,c_t,h_t), F_hat_i^0(t)=B_i(e,c_t,h_t)(2);
[0084] Where F_hat_a^0(t) represents the acoustic normal response; F_hat_i^0(t) represents the infrared normal response; B_a represents the acoustic normal response model; B_i represents the infrared normal thermal field model; e represents the device or component identifier; c_t represents the current operating condition vector; and h_t represents the device's historical status or historical normal baseline.
[0085] In some embodiments, the normal response model can be established using historical normal samples from the same device, normal samples from the same model of device, regression models, state-space models, temporal neural networks, physical-thermal models, or combinations thereof, to avoid normal load changes, ambient temperature changes, and individual device differences being mistakenly identified as abnormal. Infrared normal responses can be generated by pixel, ROI, or device component region, and can predict absolute temperature, relative ambient temperature rise, load-normalized temperature rise, temperature gradient, and hotspot area, etc.; acoustic normal responses can predict band energy, spectral envelope, harmonic amplitude, time-frequency plot, or acoustic coding vector.
[0086] In some embodiments, the online update of the normal baseline employs an anti-pollution mechanism: only samples with normal diagnostic results, modal confidence reaching the normal sample acceptance threshold, no manual alarms, and no drift detection are entered into the candidate update pool; candidate samples are used to update the normal response model after a delayed confirmation.
[0087] Subsequently, the computer equipment calculates and normalizes the difference between the actual characteristics of the voiceprint data and the normal acoustic response to obtain the acoustic residual; and calculates and normalizes the difference between the actual characteristics of the infrared data and the normal infrared response to obtain the infrared residual. The acoustic residual and the infrared residual can be calculated by the relation (3), which is shown below:
[0088] R_a(t)=Norm(F_a(t)-F_hat_a^0(t)), R_i(t)=Norm(F_i(t)-F_hat_i^0(t))(3);
[0089] Where R_a(t) and R_i(t) are the acoustic residual and infrared residual, respectively; F_a(t) and F_i(t) are the actual features of the acoustic data and the actual features of the infrared data, respectively; Norm(·) is the normalization process.
[0090] Subsequently, the computer device extracts local anomaly events from the acoustic residual R_a(t) and the infrared residual R_i(t), respectively. Specifically, the computer device extracts acoustic anomaly event a_j from the acoustic residual R_a(t) and thermal anomaly event h_k from the infrared residual R_i(t). The acoustic anomaly event a_j and the thermal anomaly event h_k can be calculated using relations (4) and (5), which are shown below:
[0091] a_j=( _j,z_j^a,b_j,l_j,s_j^a,q_j^a)(4);
[0092] in, _j represents the event time or event interval; z_j^a represents the acoustic event coding feature; b_j represents the key frequency band; l_j represents the sound source location or corresponding equipment component; s_j^a represents the event saliency; and q_j^a represents the local data quality of the event. Acoustic events can be obtained from spectral kurtosis peaks, impact envelopes, harmonic abrupt changes, periodic enhancements, sound source localization, or learning-based event detectors.
[0093] h_k=(t_k,z_k^i,r_k,grad_k,v_k,s_k^i,q_k^i)(5);
[0094] Where t_k is the thermal event time; z_k^i is the thermal event encoding feature; r_k is the ROI or component region; grad_k is the temperature gradient; v_k is the temperature rise rate; s_k^i is the event saliency; and q_k^i is the local quality. Thermal events can include the first appearance of a hot spot, hot spot area expansion, accelerated temperature rise, abrupt changes in temperature gradient, hot spot migration, and widening of interphase temperature difference, etc.
[0095] S302, input acoustic anomaly events into the prediction model to obtain prediction results, and combine the prediction results, thermal anomaly events and pre-determined hierarchical confidence levels to obtain the total mass of acoustic events and the total mass of thermal events.
[0096] Among them, the total mass of acoustic events is used to characterize the total mass of acoustic anomalies entering the cross-modal matching stage, and the total mass of thermal events is used to characterize the total mass of thermal anomalies entering the cross-modal matching stage.
[0097] In this embodiment, due to the physical time lag between acoustic anomalies and infrared thermal responses, the corresponding thermal anomaly may not have occurred when an acoustic anomaly occurs. Therefore, the computer device first predicts the acoustic anomaly using a prediction model to obtain the time delay distribution of the corresponding thermal response. The time delay distribution characterizes the probability of the corresponding thermal anomaly occurring at various future times. Subsequently, the prediction model further calculates the waiting probability based on the time delay. This waiting probability characterizes the probability that the corresponding thermal anomaly may still occur after the current time. If the waiting probability is high, it indicates that the corresponding thermal anomaly may still occur later, and the acoustic anomaly should enter a pending association state rather than be directly determined as an anomaly; if the waiting probability is low, it indicates that the corresponding thermal anomaly will not occur later, and the acoustic anomaly can be determined as an acoustic single-mode anomaly.
[0098] Subsequently, after obtaining the time delay distribution and waiting probability, the computer equipment also needs to determine the total acoustic event quality and the total thermal event quality by combining hierarchical confidence and thermal anomaly events. Specifically, the computer equipment calculates the total acoustic event quality based on the time delay distribution, waiting probability, and acoustic modal confidence; and calculates the total thermal event quality based on infrared modal confidence and thermal event local reliability. The total acoustic event quality and the total thermal event quality will serve as inputs for subsequent cross-modal matching, controlling the weights of each acoustic anomaly event and each thermal anomaly event in the matching process.
[0099] S303, the total mass of acoustic events and the total mass of thermal events are input into the matching model respectively. The matching model performs cross-modal matching according to the matching algorithm to obtain the matched event pairs, the mass of unmatched acoustic events, the mass of unmatched thermal events, and the cross-modal association confidence.
[0100] Among them, matched event pairs are used to characterize successfully matched acoustic and thermal anomaly events; unmatched acoustic event quality is used to characterize the portion of the total acoustic event quality that was not assigned to any thermal anomaly event; unmatched thermal event quality is used to characterize the portion of the total thermal event quality that was not assigned to any acoustic anomaly event. The unmatched acoustic and unmatched thermal event quality do not represent matching failures, but are used to preserve evidence of single-modal anomalies where only acoustic or thermal anomalies exist; cross-modal association credibility is used to characterize the degree of credibility of the association between acoustic and thermal anomaly events, and it is determined based on cross-modal matching support, acoustic signature modal credibility, and infrared modal credibility.
[0101] In the embodiments of this application, the total mass of acoustic events and the total mass of thermal events represent the total mass of acoustic and thermal anomalies entering the matching stage, respectively. However, the number of the two is usually not equal, and there may be real single-mode precursors or single-mode anomalies. Therefore, the matching model needs to adopt a matching algorithm that allows some event masses to remain unmatched, and to perform cross-modal matching on the total mass of acoustic events and the total mass of thermal events, so as to avoid forcibly assigning acoustic anomalies to irrelevant thermal anomalies.
[0102] Specifically, the matching model first determines the matching cost based on the time delay distribution, the feature distance between acoustic and thermal anomalies, the topological relationship between the equipment components corresponding to the acoustic and thermal anomalies, and the local reliability of the acoustic and thermal anomalies. The matching cost characterizes the difficulty of matching acoustic and thermal anomalies; a lower matching cost indicates a higher probability that the two events correspond to the same equipment anomaly. Subsequently, the matching model performs cross-modal matching on the total quality of acoustic and thermal events according to a pre-defined matching algorithm and the matching cost, obtaining the matching results. These results include matched event pairs, the quality of unmatched acoustic events, the quality of unmatched thermal events, and the cross-modal association reliability.
[0103] In some embodiments, the computer device may replace the matching algorithm in the matching model with a cross-attention mechanism. Specifically, the computer device may introduce an event-level local reliability vector, as well as a time delay bias and a device topology bias, into the cross-attention mechanism. It should be noted that the global modal reliability is a single scalar. If the same scalar is uniformly added to all attention logic values for the same query, the attention distribution will not change after Softmax normalization. Therefore, this application uses a local reliability vector that varies with the key position in the attention logic values, while using the global modal reliability for branch gating. Specifically, cross-modal attention can be represented by relation (6), which is shown below:
[0104] A_ai=softmax(Q_aK_i^T / sqrt(d)+lambda_r·1·log(r_i+eps)^T+lambda_tB_Delta+lambda_gB_G), F_ai=A_aiV_i(6);
[0105] Where A_ai represents the cross-modal attention matrix; Q_a represents the query matrix for acoustic anomalies; K_i represents the key matrix for thermal anomalies; sqrt(d) represents the scaling factor; lambda_r, lambda_t, and lambda_g represent the weight parameters; 1 represents a one vector; r_i represents the local reliability vector for thermal anomalies; eps represents the minimum constant to prevent taking the logarithm of zero; B_Delta represents the time-delay compatible bias matrix; B_G represents the device component topology bias matrix; F_ai represents the cross-modal attention feature; V_i represents the value matrix for thermal anomalies; and softmax(·) represents the normalized exponential function.
[0106] It should be noted that the above formula (6) is a cross-attention method with acoustic anomalies as the query and thermal anomalies as the key and value. In practical applications, reverse attention can also be used, that is, cross-attention with thermal anomalies as the query and acoustic anomalies as the key and value, and its calculation method is the same as formula (6). The attention features of the two directions are used together for subsequent fusion, which can mutually verify the correlation between acoustic anomalies and thermal anomalies.
[0107] After obtaining the cross-attention feature F_ai, the computer device can further perform gating processing on the cross-modal association confidence level C_ai. Specifically, when the cross-modal association confidence level C_ai is high, it indicates that the association between acoustic anomalies and thermal anomalies is relatively reliable, and more cross-modal attention features are retained; when the cross-modal association confidence level C_ai is low, it indicates that the association between acoustic anomalies and thermal anomalies is not very reliable, weakening the cross-modal attention features.
[0108] The cross-modal matching method based on cross-attention described above is applicable to systems that have already adopted the Transformer structure for feature processing. For such systems, it is not necessary to replace the matching algorithm with an unbalanced optimal transport to achieve cross-modal matching of acoustic anomalies and thermal anomalies.
[0109] S304. The matched event pairs, the quality of the unmatched acoustic events, the quality of the unmatched thermal events, and the cross-modal correlation confidence are input into the replacement model. The replacement model replaces the acoustic residuals with the normal acoustic response under the current operating conditions and the infrared residuals with the normal infrared response under the current operating conditions. The changes in the diagnostic conclusions before and after the replacement are calculated to obtain the voiceprint evidence contribution and the infrared evidence contribution. The matched event pairs, the quality of the unmatched acoustic events, the quality of the unmatched thermal events, the cross-modal correlation confidence, the hierarchical confidence, the voiceprint evidence contribution, and the infrared evidence contribution are fused to obtain the abnormal evolution state and the diagnostic results.
[0110] Among them, the voiceprint evidence contribution is used to characterize the actual driving effect of abnormal evidence in the voiceprint modality on abnormal conclusions, and the infrared evidence contribution is used to characterize the actual driving effect of abnormal evidence in the infrared modality on abnormal conclusions.
[0111] In this embodiment, after receiving the matched event pairs, unmatched acoustic event quality, unmatched thermal event quality, and cross-modal association confidence from the matching model, the replacement model needs to further determine the contribution of acoustic fingerprint evidence and infrared evidence to verify the actual driving role of each modality in the anomalous conclusion. High quality of modal data does not necessarily mean that the anomalous evidence of that modality has an actual contribution to the final diagnostic conclusion; therefore, counterfactual substitution is required for verification.
[0112] Specifically, the replacement model replaces the acoustic residuals with the normal acoustic response under the current operating conditions and the infrared residuals with the normal infrared response under the current operating conditions, respectively, and calculates the change in diagnostic conclusions before and after the replacement. A large change in the diagnostic conclusions after replacement indicates a strong driving effect of the anomalous evidence for that mode on the anomalous conclusions, resulting in a high evidence contribution; conversely, a small change indicates a weak driving effect of the anomalous evidence for that mode on the anomalous conclusions, resulting in a low evidence contribution. After obtaining the acoustic and infrared evidence contributions, the replacement model fuses the matched event pairs, the quality of unmatched acoustic events, the quality of unmatched thermal events, and the cross-modal association confidence, along with the acoustic and infrared evidence contributions and the hierarchical confidence, to obtain fused features. These fused features are then subjected to state discrimination processing to obtain the anomalous evolution state and diagnostic results. The anomalous evolution state characterizes the current anomalous stage of the target power equipment.
[0113] In an exemplary embodiment, the above-mentioned S303, "performing cross-modal matching according to the matching algorithm to obtain matched event pairs, unmatched acoustic event quality, unmatched thermal event quality, and cross-modal association confidence," is as follows: Figure 4 As shown, it includes:
[0114] S401, determine the matching cost based on the time delay distribution, the characteristic distance between acoustic anomalies and thermal anomalies, the topological relationship between the corresponding equipment components of acoustic anomalies and thermal anomalies, and the local reliability of acoustic anomalies and thermal anomalies.
[0115] In this embodiment, the computer device determines the matching cost based on the time delay distribution, the characteristic distance between acoustic anomalies and thermal anomalies, the topological relationship between the device components corresponding to the acoustic anomalies and the thermal anomalies, and the local reliability of the acoustic and thermal anomalies. Specifically, the matching cost can be represented by equation (7), which is shown below:
[0116] C_jk=lambda_t[-ln(p_j( _jk)+eps)]+lambda_f[1-sim(P_az_j^a,P_iz_k^i)]+lambda_gd_G(l_j,r_k)+lambda_q[(1-r_aj)+(1-r_ik)](7);
[0117] Where C_jk represents the matching cost between the j-th acoustic anomaly and the k-th thermal anomaly; lambda_t, lambda_f, lambda_g, and lambda_q are weight parameters; p_j( (_jk) represents the time delay distribution of the thermal response corresponding to the j-th acoustic anomaly event, where the time delay is... The probability of _jk; _jk represents the time difference between the j-th acoustic anomaly and the k-th thermal anomaly; eps represents the minimum constant to prevent taking the logarithm of zero; sim(·) represents the similarity function; P_a represents the acoustic feature projection matrix; P_i represents the thermal feature projection matrix; z_j^a represents the acoustic event coding feature of the j-th acoustic anomaly; z_k^i represents the thermal event coding feature of the k-th thermal anomaly; d_G(·) represents the topological distance function between the equipment and components; l_j represents the sound source location or corresponding equipment component of the j-th acoustic anomaly; r_k represents the ROI or component region of the k-th thermal anomaly; r_aj represents the local reliability of the j-th acoustic anomaly; r_ik represents the local reliability of the k-th thermal anomaly.
[0118] It is important to note that this matching cost allows the system to distinguish between events that are "timely but component-independent" and events that are "timely but consistent with the thermal inertia of the equipment." If the sound source location area and the hotspot area are located on the same component or have a clear structural / thermal conduction relationship, the topological distance is small; if they are located on unrelated components, the matching cost increases.
[0119] S402, perform cross-modal matching on the total quality of acoustic events and the total quality of thermal events based on the matching algorithm and matching cost, and obtain the matched event pairs, the quality of unmatched acoustic events, the quality of unmatched thermal events, and the cross-modal association confidence.
[0120] In this embodiment, after obtaining the matching cost, the computer device can use the matching cost as the target to match the total mass of acoustic events and the total mass of thermal events using a matching algorithm, thereby obtaining a matching matrix between acoustic anomalies and thermal anomalies. Specifically, the matching matrix can be represented by relation (8), which is shown below:
[0121] Pi*=argmin_{Pi>=0}[ _jkPi_jkC_jk+eps_ot _jkPi_jk(lnPi_jk-1)+rho_aKL(Pi·1||mu_a)+rho_iKL(Pi^T·1||mu_i)](8);
[0122] Where Pi* represents the optimal matching matrix from acoustic anomaly to thermal anomaly; Pi represents the matching matrix from acoustic anomaly to thermal anomaly; Pi_jk represents the element value between the j-th acoustic anomaly and the k-th thermal anomaly in the matching matrix; C_jk represents the matching cost between the j-th acoustic anomaly and the k-th thermal anomaly; eps_ot represents the entropy regularization parameter; rho_a and rho_i represent the marginal deviation penalty coefficients; KL(·||·) represents the KL divergence; Pi·1 represents the marginal distribution of the summation of the matching matrix over the direction of the thermal anomaly; Pi^T·1 represents the marginal distribution of the summation of the matching matrix over the direction of the acoustic anomaly; mu_a represents the total mass of the acoustic event; and mu_i represents the total mass of the thermal event.
[0123] It is important to note that this matching algorithm allows the total mass of acoustic events to be less than or equal to the total mass of thermal events, and allows some event masses to remain unmatched. Therefore, masses without corresponding events can be explicitly preserved. For example, this matching algorithm can be an algorithm for solving unbalanced optimal transport problems.
[0124] Secondly, the computer device determines the matched event pairs, the quality of unmatched acoustic events, the quality of unmatched thermal events, and the cross-modal matching support based on the matching matrix. Specifically, the quality of unmatched acoustic events, the quality of unmatched thermal events, and the cross-modal matching support can be represented by relations (9) and (10), respectively, as shown below:
[0125] u_a=[mu_a-Pi·1]_+, u_i=[mu_i-(Pi)^T·1]_+ (9);
[0126] Where u_a represents the quality of unmatched acoustic events; u_i represents the quality of unmatched thermal events; mu_a represents the total quality of acoustic events; mu_i represents the total quality of thermal events; Pi* represents the optimal matching matrix; [·]_+ represents the positive value operation, that is, when the value in the parentheses is negative, it is zero.
[0127] M_ai=[Σ_jkPi_jkexp(-C_jk)] / [Σ_jkPi_jk+eps] (10);
[0128] Where M_ai represents the cross-modal matching support; Pi*_jk represents the element value between the j-th acoustic anomaly and the k-th thermal anomaly in the optimal matching matrix; C_jk represents the matching cost between the j-th acoustic anomaly and the k-th thermal anomaly; and eps represents a minimal constant to prevent the denominator from being zero.
[0129] Specifically, the computer device determines the acoustic event quality corresponding to an unmatched acoustic anomaly event as the unmatched acoustic event quality u_a, the thermal event quality corresponding to an unmatched thermal anomaly event as the unmatched thermal event quality u_i, and acoustic anomalies and thermal anomalies with matching quality higher than a preset threshold as matched event pairs.
[0130] Then, the computer device determines the cross-modal association confidence based on the cross-modal matching support, voiceprint modal confidence, and infrared modal confidence. Specifically, the cross-modal association confidence can be represented by relation (11), which is shown below:
[0131] C_ai=sqrt(C_aC_i)·M_ai(11);
[0132] Where C_ai represents cross-modal association confidence; C_a represents voiceprint modal confidence; C_i represents infrared modal confidence; M_ai represents cross-modal matching support; and sqrt(·) represents the square root function.
[0133] It is important to note that C_a and C_i describe whether the modal data itself is reliable, while C_ai describes whether the correlation between the current acoustic anomaly and the thermal anomaly is reliable. These two types of reliability have different meanings.
[0134] It is important to note that when the acoustic modal confidence level C_a is higher than the preset confidence level threshold and the unmatched acoustic event quality u_a is higher than the preset quality threshold, it indicates that the thermal anomaly event corresponding to the acoustic anomaly event has not yet occurred. The computer device uses the waiting probability to determine whether the acoustic anomaly event is in the acoustic precursor stage. When the infrared modal confidence level C_i is higher than the preset confidence level threshold and the unmatched thermal event quality u_i is higher than the preset quality threshold, it indicates that the acoustic anomaly event corresponding to the thermal anomaly event does not exist. The computer device identifies the thermal anomaly event as evidence of a single thermal modal anomaly. When both the unmatched acoustic event quality u_a and the unmatched thermal event quality u_i are higher than the preset quality threshold, and the equipment component corresponding to the unmatched acoustic event is different from the equipment component corresponding to the unmatched thermal event, it indicates that the target power equipment has multiple independent anomalies. The computer device identifies this as suspected evidence of a composite anomaly.
[0135] In an exemplary embodiment, the above-mentioned S302, "inputting acoustic anomaly events into the prediction model, obtaining prediction results, and combining the prediction results, thermal anomaly events, and pre-determined hierarchical confidence levels to obtain the total acoustic event mass and the total thermal event mass," is as follows: Figure 5 As shown, it includes:
[0136] S501, input the acoustic anomaly event into the prediction model for prediction, and obtain the time delay distribution and waiting probability corresponding to the acoustic anomaly event.
[0137] The waiting probability is used to characterize the probability that an acoustic anomaly will give rise to a corresponding thermal anomaly after the current moment.
[0138] In this embodiment, the computer device inputs the acoustic anomaly event a_j into the prediction model. The prediction model includes a conditional time-delay prediction model g_theta, which is used to predict the time delay distribution of the acoustic anomaly event under the current equipment and operating conditions, corresponding to the thermal anomaly. Unlike selecting infrared frames based on a fixed threshold, the time delay distribution is jointly determined by the acoustic event characteristics, equipment attributes, and current operating conditions. Specifically, the time delay distribution can be represented by the relation (12), which is shown below:
[0139] p_j( )=Softmax(g_theta(z_j^a,e,F_c))_ , ∈ (12);
[0140] Among them, p_j( ) represents the time delay distribution of the thermal response corresponding to the j-th acoustic anomaly event; Indicates candidate time delay; represents the candidate time delay set; g_theta represents the conditional time delay prediction model; z_j^a represents the acoustic event encoding feature; e represents the device or component identifier; F_c represents the operating condition feature; Softmax(·) represents the normalized exponential function.
[0141] It is important to note that the time delay can be represented by discrete bins, or by log-normal, Gamma, mixed distributions, quantile regression, or conditional density networks. For anomalies such as mechanical loosening or friction, the time delay distribution can have a long tail; for rapid heating due to poor contact, the time delay distribution can be concentrated in a shorter interval.
[0142] When no thermal anomaly event has been found for a certain acoustic event, the computer device calculates the probability that the acoustic event may still produce a thermal response after the current moment, i.e., the waiting probability. Specifically, the waiting probability can be represented by the relation (13), which is shown below:
[0143] P_j^wait(t) = _{ ∈ , >t- _j}p_j( (13);
[0144] Where P_j^wait(t) represents the waiting probability; t represents the current time. _j represents the event time of the acoustic anomaly event.
[0145] It is important to note that if the waiting probability is higher than the preset waiting threshold, it indicates that the thermal anomaly corresponding to the acoustic anomaly may still occur later. The computer device adds the acoustic anomaly to the queue to be associated and outputs the acoustic precursor to be associated status. If the waiting probability is not higher than the waiting threshold, or if the corresponding thermal anomaly does not occur after exceeding the high quantile of the time delay distribution, it indicates that the thermal anomaly corresponding to the acoustic anomaly will not occur later. The computer device judges the acoustic anomaly as an acoustic single-mode anomaly or a modal anomaly based on the anomaly type and other evidence.
[0146] S502 determines the total mass of acoustic events based on time delay distribution, waiting probability, and hierarchical credibility.
[0147] The layered reliability consists of two levels: the modal layer and the event layer. Specifically, the modal layer includes acoustic signature modal reliability C_a and infrared modal reliability C_i, used to characterize the reliability of the acoustic signature data and infrared data themselves, respectively. The event layer includes acoustic event local reliability r_aj and thermal event local reliability r_ik, used to characterize the local quality of each acoustic event and each thermal anomaly event within the same modality, respectively. It is important to note that the acoustic signature modal reliability C_a and the infrared modal reliability C_i do not decrease simply because another modality is temporarily free of anomalies.
[0148] In this embodiment, the computer device first determines the voiceprint modal confidence C_a based on the quality index of the voiceprint data, and determines the infrared modal confidence C_i based on the quality index of the infrared data. Specifically, the modal confidence can be represented by the relation (14), which is shown below:
[0149] C_m=sigmoid(b_m+w_m1Q_m+w_m2G_m+w_m3S_m^pert+w_m4D_m-w_m5U_m)(14);
[0150] Where m∈{a,i}, m=a represents the voiceprint mode, and m=i represents the infrared mode; C_m represents the mode confidence; b_m represents the bias term; Q_m represents the overall data quality of the mode; G_m represents the working condition adaptability; S_m^pert represents the prediction stability under reasonable sensor perturbation; D_m represents the degree to which the current sample belongs to the training / calibration distribution; U_m represents the prediction uncertainty of the single-mode model; b_m and w_m1~w_m5 are learnable or configurable parameters; sigmoid(·) represents the sigmoid activation function.
[0151] In some embodiments, Q_a can be obtained from signal-to-noise ratio, effective sampling rate, packet loss rate, amplitude saturation ratio, etc.; Q_i can be obtained from temperature matrix integrity, ROI integrity, sharpness, invalid pixel rate, etc.
[0152] Secondly, the computer equipment determines the local reliability r_aj of the acoustic anomaly based on the local signal-to-noise ratio, array positioning stability, and event-level prediction uncertainty of the acoustic anomaly; and determines the local reliability r_ik of the thermal anomaly based on the ROI occlusion rate, local temperature matrix efficiency, hotspot stability, and local prediction uncertainty of the thermal anomaly.
[0153] Finally, the computer device determines the total mass of acoustic events based on the modal reliability of the acoustic signature, the local reliability of the acoustic events, and the event saliency of the acoustic anomalies. Specifically, the total mass of acoustic events can be represented by relation (15), which is shown below:
[0154] mu_aj=C_a·(s_j^a·r_aj) / (Σ_ls_l^a·r_al+eps)(15);
[0155] Where mu_aj represents the total acoustic event quality of the j-th acoustic anomaly; C_a represents the speaker modal reliability; s_j^a represents the event significance of the j-th acoustic anomaly; r_aj represents the local reliability of the j-th acoustic anomaly; s_l^a represents the event significance of the l-th acoustic anomaly; r_al represents the local reliability of the l-th acoustic anomaly; and eps represents a minimal constant to prevent the denominator from being zero.
[0156] As a result, the total event quality of low-confidence modes entering the matching process is automatically reduced; the quality of events that are locally occluded, contaminated by noise, or have high uncertainty within the same mode is also reduced.
[0157] S503, determine the total mass of the thermal event based on the infrared modal reliability of the layered reliability and the local reliability of the thermal event of the thermal anomaly.
[0158] In this embodiment, the computer device determines the total mass of a thermal event based on infrared modal reliability, local reliability of the thermal event, and event saliency of the thermal anomaly. Specifically, the total mass of the thermal event can be represented by equation (16), which is shown below:
[0159] mu_ik=C_i·(s_k^i·r_ik) / (Σ_ls_l^i·r_il+eps)(16);
[0160] Where mu_ik represents the total thermal event mass of the k-th thermal anomaly; C_i represents the infrared modal reliability; s_k^i represents the event significance of the k-th thermal anomaly; r_ik represents the local reliability of the k-th thermal anomaly; s_l^i represents the event significance of the l-th thermal anomaly; r_il represents the local reliability of the l-th thermal anomaly; and eps represents the minimum constant to prevent the denominator from being zero.
[0161] In an exemplary embodiment, the above-mentioned step S304, "fusing the matched event pairs, the quality of the unmatched acoustic events, the quality of the unmatched thermal events, the cross-modal association confidence, the hierarchical confidence, the acoustic evidence contribution, and the infrared evidence contribution to obtain the abnormal evolution state and diagnostic results," is as follows: Figure 6 As shown, it includes:
[0162] S601, based on the matched event pairs, perform weighted aggregation on the matched acoustic anomaly events and thermal anomaly events to obtain the matched event pair features.
[0163] In this embodiment, the computer device performs weighted aggregation of the matched acoustic anomaly events and thermal anomaly events based on the matching quality of each event pair in the matched event pair to obtain the matching event pair features. Specifically, the matching event pair features can be represented by relation (17), which is shown below:
[0164] F_pair=[ _jkPi_jk·phi(z_j^a,z_k^i, _jk,d_G)] / [ _jkPi_jk+eps](17;
[0165] Where F_pair represents the matching event pair feature; Pi*_jk represents the element value between the j-th acoustic anomaly event and the k-th thermal anomaly event in the optimal matching matrix; phi(·) represents the feature projection function; z_j^a represents the acoustic event encoding feature of the j-th acoustic anomaly event; and z_k^i represents the thermal event encoding feature of the k-th thermal anomaly event. _jk represents the time difference between the j-th acoustic anomaly and the k-th thermal anomaly; d_G represents the topological distance between the equipment component corresponding to the acoustic anomaly and the equipment component corresponding to the thermal anomaly; eps represents a minimal constant to prevent the denominator from being zero.
[0166] S602, unmatched acoustic anomaly events and unmatched thermal anomaly events are weighted and aggregated according to the quality of unmatched acoustic events and the quality of unmatched thermal events respectively to obtain unmatched single-modal features.
[0167] In this embodiment, the computer device performs weighted aggregation of unmatched acoustic anomaly events based on the quality of unmatched acoustic events, and performs weighted aggregation of unmatched thermal anomaly events based on the quality of unmatched thermal events. The aggregation results are then concatenated to obtain the unmatched single-modal features. Specifically, the unmatched single-modal features can be represented by relation (18), which is shown below:
[0168] F_single=concat( _ju_a,jpsi_a(z_j^a), _ku_i,kpsi_i(z_k^i))(18);
[0169] Where F_single represents the unmatched single-modal feature; concat(·) represents the concatenation function; u_a,j represents the unmatched acoustic event quality of the j-th unmatched acoustic anomaly; psi_a(·) represents the acoustic feature projection function; z_j^a represents the acoustic event encoding feature of the j-th unmatched acoustic anomaly; u_i,k represents the unmatched thermal event quality of the k-th unmatched thermal anomaly; psi_i(·) represents the thermal feature projection function; and z_k^i represents the thermal event encoding feature of the k-th unmatched thermal anomaly.
[0170] After obtaining the features of the matching event pairs and the features of the unmatched single modalities, the computer equipment also needs to determine the contribution of voiceprint evidence and infrared evidence to verify the actual driving role of each modality in the conclusion of anomalies.
[0171] It should be noted that high quality of modal data does not necessarily mean that the anomalous evidence of that modality actually contributes to the final diagnostic conclusion. To verify whether a certain modality truly drives the anomalous conclusion, the computer equipment constructs a normal prototype that matches the operating conditions after initial fusion. The normal prototype is then applied to a single modality, while keeping the other modality, operating conditions, and equipment component information unchanged, thereby determining the contribution of acoustic signature evidence and infrared evidence.
[0172] Specifically, the computer device determines the preliminary fusion anomaly score s based on the acoustic residual, infrared residual, and operating condition characteristics; constructs a normal acoustic prototype R_a^0 and a normal thermal prototype R_i^0 under the current operating condition, and replaces the acoustic residual with the normal acoustic prototype and the infrared residual with the normal thermal prototype, respectively, while keeping the other mode, operating condition, and equipment component information unchanged, to obtain the fusion anomaly score after replacement; and determines the acoustic evidence contribution and infrared evidence contribution based on the change in the fusion anomaly score before and after replacement. Specifically, the acoustic evidence contribution and infrared evidence contribution can be represented by the relations (19) and (20), which are shown below:
[0173] s=f_pre(R_a,R_i,F_c), s_a^cf=f_pre(R_a^0,R_i,F_c), s_i^cf=f_pre(R_a,R_i^0,F_c) (19);
[0174] I_a^cf=clip((s-s_a^cf) / (|s|+eps),0,1), I_i^cf=clip((s-s_i^cf) / (|s|+eps),0,1) (20);
[0175] Where s represents the initial fusion anomaly score; s_a^cf represents the fusion anomaly score after replacing the acoustic residual with the normal acoustic prototype; s_i^cf represents the fusion anomaly score after replacing the infrared residual with the normal thermal prototype; f_pre(·) represents the fusion function; R_a represents the acoustic residual; R_i represents the infrared residual; R_a^0 represents the normal acoustic prototype under the current device and current operating conditions; R_i^0 represents the normal thermal prototype under the current device and current operating conditions; F_c represents the operating condition characteristics; I_a^cf represents the acoustic signature contribution; I_i^cf represents the infrared signature contribution; clip(·) represents the truncation function; and eps represents a minimal constant to prevent the denominator from being zero.
[0176] It should be noted that counterfactual contributions are used for interpreting anomalous evidence and modulating the final fusion branch, but are not directly equivalent to the data credibility of the modality itself, thus avoiding the misinterpretation of "there is currently no anomalous evidence" as "the data of this modality is unreliable".
[0177] S603 concatenates the matching event pair features, unmatched single-modal features, hierarchical credibility, cross-modal association credibility, voiceprint evidence contribution, and infrared evidence contribution to obtain fused features.
[0178] In this embodiment, the computer device concatenates the matching event pair features, unmatched single-modal features, hierarchical confidence, cross-modal association confidence, voiceprint evidence contribution, infrared evidence contribution, and working condition data to obtain fused features. Specifically, the fused features can be represented by relation (21), which is shown below:
[0179] F=concat(F_pair,F_single,C_aF_a^global,C_iF_i^global,C_ai,I_a^cf,I_i^cf,F_c)(21);
[0180] Where F represents the fusion feature; F_pair represents the matched event pair feature; F_single represents the unmatched single-modal feature; C_a represents the voiceprint modality confidence; F_a^global represents the voiceprint global feature; C_i represents the infrared modality confidence; F_i^global represents the infrared global feature; C_ai represents the cross-modal association confidence; I_a^cf represents the voiceprint evidence contribution; I_i^cf represents the infrared evidence contribution; F_c represents the working condition feature; and concat(·) represents the concatenation function.
[0181] It should be noted that the fusion feature F retains matched cross-modal evidence, unmatched single-modal evidence, modality credibility, cross-modal association credibility, counterfactual evidence contribution, and operational information. Therefore, it will not lose precursors or compound anomalies due to forced fusion.
[0182] S604 performs state discrimination processing on the fusion features to obtain abnormal evolution states and diagnostic results.
[0183] In this embodiment, the computer device performs state discrimination processing on the fusion feature F to obtain the abnormal evolution state and the diagnostic result. Specifically, the computer device obtains the acoustic single-mode posterior probability p_a corresponding to the acoustic abnormal event and the infrared single-mode posterior probability p_i corresponding to the thermal abnormal event, and calculates the divergence degree D_JS between the two. Subsequently, the computer device performs state discrimination processing based on the fusion feature F, the acoustic single-mode posterior probability p_a, the infrared single-mode posterior probability p_i, the divergence degree D_JS, the quality u_a of the unmatched acoustic event, the quality u_i of the unmatched thermal event, the waiting probability P^wait, and the abnormal evolution state z_{t-1} of the previous moment to obtain the abnormal evolution state z_t of the current moment and the diagnostic result. Specifically, the abnormal evolution state of the current moment can be represented by the relation (22), which is shown below:
[0184] z_t=argmax_zP(z|F,u_a,u_i,P^wait,D_JS,z_{t-1})(22);
[0185] Where z_t represents the current anomalous evolution state; argmax_z represents the anomalous evolution state that maximizes the probability; P(·) represents the conditional probability; F represents the fusion feature; u_a represents the quality of the unmatched acoustic event; u_i represents the quality of the unmatched thermal event; P^wait represents the waiting probability; D_JS represents the degree of divergence; and z_{t-1} represents the anomalous evolution state of the previous time step.
[0186] Abnormal evolution states include at least one of the following: insufficient data quality requiring review, i.e., both the voiceprint modal confidence C_a and the infrared modal confidence C_i are below the low confidence threshold, or the quality of valid events is insufficient; acoustic precursors awaiting association, i.e., the voiceprint modal confidence C_a is high, there are significant unmatched acoustic events, and the waiting probability P^wait is high, indicating that the expected thermal response may still occur later; thermal single-modal anomaly, i.e., the infrared modal confidence C_i is high and there are significant unmatched thermal events, while the voiceprint data itself is reliable but there is no corresponding acoustic anomaly. Thermal anomaly formation refers to the occurrence of a subsequent thermal event within the prediction time lag interval after the existence of a previously uncorrelated acoustic event, resulting in a valid match. Acoustic-thermal synergistic anomaly refers to the high cross-modal matching support M_ai and cross-modal association confidence C_ai, where acoustic and thermal events corroborate each other at the time lag, component topology, and feature levels. Suspected composite / conflict anomaly refers to the high acoustic modal confidence C_a and infrared modal confidence C_i, but a high divergence D_JS, and the acoustic and thermal events pointing to different components or low cross-modal matching support.
[0187] The diagnostic results should include at least a Level 1 anomaly state, and may further output anomaly type tendency and anomaly evolution stage. Level 1 states may include normal, acoustic precursors pending correlation, single-mode acoustic anomaly, single-mode thermal anomaly, thermal anomaly formation, acoustic-thermal synergistic anomaly, suspected composite / conflict anomaly, and insufficient data quality requiring review. Anomaly type tendency may include mechanical vibration, friction / impact, mechanical loosening, suspected partial discharge, poor contact / heating at connection points, heat dissipation / cooling anomaly, load-related temperature rise anomaly, and unknown anomaly.
[0188] In some embodiments, in addition to outputting abnormal evolution states and diagnostic results, the computer device may also output interpretable output, which includes key time periods of the voiceprint, key frequency bands, sound source location, voiceprint modal credibility; infrared hotspot regions, relative temperature differences, hotspot areas, temperature rise rates, infrared modal credibility; acoustic-thermal matching event pairs, measured time delays, predicted time delay intervals, cross-modal association credibility, unmatched events, counterfactual evidence contributions, and suggested review items.
[0189] In some embodiments, state discrimination processing can be implemented using finite state machines, hidden Markov models, conditional random fields, temporal neural networks, or a hybrid approach of rules and models. If an acoustic precursor appears with a corresponding hotspot within the prediction time delay, it can be classified as a thermal anomaly or a combined acoustic-thermal anomaly; if no hotspot appears beyond the high quantile time delay, it is determined to be a long-term acoustic single-mode anomaly or enters the verification stage based on the candidate anomaly type.
[0190] In an exemplary embodiment, the diagnostic model can be trained in the following ways: the normal response model is trained preferentially using normal samples; the conditional time delay prediction model in the prediction model is trained under supervision or weak supervision using the time difference between acoustic events and thermal events in historical alarms, maintenance records, and continuous inspection data; the matching model and the replacement model are trained using supervised, semi-supervised, or self-supervised methods. Specifically, the training loss of the diagnostic model can be represented by relation (23), which is shown below:
[0191] L=L_base+lambda_1L_lag+lambda_2L_match+lambda_3L_cls+lambda_4L_state+lambda_5L_cal(23);
[0192] Where L represents the total training loss of the diagnostic model; L_base represents the normal response prediction loss; L_lag represents the time-delay distribution prediction loss; L_match represents the cross-modal matching constraint loss or the optimal transmission target of relation (11); L_cls represents the abnormal state / type classification loss; L_state represents the abnormal evolution stage loss; L_cal represents the modal confidence calibration loss; lambda_1~lambda_5 represent the weight parameters of each loss term. Each loss term can be selected and used in part according to the data conditions.
[0193] During training, modal credibility can be obtained through rule functions, logistic regression, gradient boosting models, or neural networks. Preferably, monotonic constraints are applied to the modal credibility calculation model to ensure that credibility does not decrease as data quality, operational adaptability, perturbation stability, and intra-domain extent increase, and that credibility does not increase as model uncertainty increases.
[0194] After the diagnostic model is trained and deployed in a real-time online monitoring scenario, the computer equipment maintains a device-level state cache according to a preset diagnostic cycle. This cache includes the most recent acoustic anomaly events, the most recent thermal anomaly events, the prediction time delay distribution of each acoustic anomaly event, the waiting probability, the anomaly evolution state at the previous moment, and the confidence statistics. When the frame rate of the infrared data is lower than the sampling rate of the acoustic signature data, the acoustic anomaly event can first enter the waiting association queue and continuously match it with subsequent thermal anomaly events within the prediction time delay range.
[0195] During online monitoring, computer equipment monitors acoustic signature spectrum baseline drift, infrared temperature distribution drift, sensor gain changes, target area positioning drift, and operating condition distribution drift. If a significant drift is detected, the intra-domain degree D_m of the corresponding mode is reduced, thereby reducing the mode confidence and entering a verification mode. Unverified anomalous samples and high-confidence conflicting samples must not be directly updated in the normal response model to prevent anomalies from contaminating the normal baseline.
[0196] The methods described in each of the above steps have been described in the foregoing embodiments. For details, please refer to the foregoing descriptions. They will not be repeated here.
[0197] The above embodiments are explained and illustrated by some examples below, which do not limit the technical solution.
[0198] Example 1 is applied to a scenario where a delayed thermal response occurs after an acoustic precursor.
[0199] The station server receives 120 seconds of continuous acoustic signature signals from a power connection point, along with an infrared temperature matrix every 5 seconds, load current, and ambient temperature. The current load is 0.78 of the rated load, and the ambient temperature is 31.2℃. The normal response model predicts that the normalized energy of the key acoustic signature frequency band under this condition is 0.36, and the normal relative temperature rise of the target area is 8.1. .
[0200] Acoustic anomaly event a_1 was detected at 12.4s, with a residual significance of 0.82 and a local quality q_1^a of 0.91. The overall signal-to-noise ratio, effective sampling rate, and operational adaptability of the acoustic signature modality were all good, and the calculated confidence level of the acoustic signature modality C_a = 0.88. The conditional time delay prediction model in the prediction unit gives the probabilities of discrete time delay centers {15, 35, 55, 75, 105}s as {0.05, 0.15, 0.45, 0.25, 0.10}, respectively. No thermal anomaly event occurred at 35.0s. At this point, 22.6s had passed since the acoustic anomaly event, and the waiting probability P^wait was still relatively high. Therefore, the computer device kept a_1 in the waiting-to-associate queue and output "Acoustic precursor waiting to be associated," without interpreting the absence of a hotspot in the current infrared light as an unreliable acoustic signature.
[0201] Thermal anomaly event h_1 was detected at 62.4 s, with a measured time lag of approximately 50.0 s compared to a_1, falling within the high-probability interval of the time lag distribution. h_1 is located at the same connection point, the target area integrity is 0.96, and the relative temperature rise is 8.6. Rising to 13.8 The hotspot area increases. The infrared modal confidence level C_i = 0.90. Based on the time delay distribution, feature distance, equipment component topology, and local reliability, the matching model determines that the time delay cost and topology cost in the matching cost are relatively small. An effective matching algorithm that allows some events to remain unmatched is adopted to establish a_1-h_1. The cross-modal matching support increases, and the anomaly evolution state changes from "acoustic precursors to be associated" to "thermal anomaly formation". If subsequent acoustic anomaly events and hotspots continue to increase, it will further change to "acoustic-thermal coordinated anomaly".
[0202] In this embodiment, after replacing the acoustic residual with a normal acoustic prototype in the replacement model, the initial fusion anomaly score was significantly reduced; similarly, after replacing the infrared residual with a normal thermal field prototype, the anomaly score was also reduced, indicating that both acoustic evidence and infrared evidence contribute substantially to the final conclusion. This process reflects the explicit modeling of the "acoustic-then-thermal" evolution process in this application.
[0203] Example 2 is applied to a scenario of high-confidence thermal single-mode anomaly.
[0204] Localized hotspots appeared in the infrared thermal field at a busbar connection point, with a target area integrity of 0.96. The relative temperature difference and temperature rise rate of the hotspots significantly deviated from the normal thermal field response under the current load; the infrared modal confidence level Ci = 0.92. Simultaneously, the acoustic fingerprint acquisition quality was good, with an acoustic fingerprint modal confidence level Ca = 0.87, but no significant anomalous events were formed in the acoustic residuals. Since "no acoustic fingerprint anomaly" is not equivalent to "low-quality acoustic fingerprint data," the computer equipment retained a higher Ca, while the thermal anomaly event resulted in a large unmatched thermal event quality ui. The final output was "thermal single-modal anomaly," rather than explaining the modal differences by reducing the acoustic fingerprint confidence level.
[0205] Example 3 is applied to a scenario where two high-confidence modes point to different components.
[0206] The acoustic array detected periodic impact events in the cooling fan bearing area, with high acoustic modal reliability. Simultaneously, the infrared thermal field detected persistent hot spots at the sleeve connection, also with high infrared modal reliability. The topological distance between the equipment components corresponding to the two events is large, and the topological cost in the matching cost constructed by the matching model prevents them from being forcibly matched. A matching algorithm that allows some event quality to remain unmatched retains higher unmatched acoustic event quality u_a and unmatched thermal event quality u_i. The replacement model, combined with a high modal divergence degree D_JS, outputs "Suspected Composite / Conflict Anomaly," providing explanations for the mechanical anomaly in the fan bearing and the thermal anomaly in the sleeve connection, respectively, avoiding averaging two real anomalies into a single medium-risk conclusion.
[0207] Example 4 is applied to scenarios where low-quality data is rejected.
[0208] Strong winds and heavy rain resulted in a low signal-to-noise ratio for acoustic signatures, and the infrared target area was obscured, significantly reducing the proportion of effective temperature pixels. Although a single-modal anomaly classifier might give a high anomaly probability, the acoustic signature data quality Q_a, infrared data quality Q_i, and intra-domain degree D_m were low, and the perturbation stability was also poor. Therefore, the calculated acoustic signature modal confidence C_a and infrared modal confidence C_i were both below the low confidence threshold. The computer output "Data quality insufficient, needs to be reviewed" and prompted the user to re-collect acoustic signatures, adjust the infrared perspective, or wait for weather conditions to improve before retesting, without forcibly outputting "normal" or "abnormal."
[0209] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.
[0210] Based on the same inventive concept, this application also provides an anomaly diagnosis device for substation power equipment to implement the above-described anomaly diagnosis method for substation power equipment. The solution provided by this device is similar to the implementation described in the above method. Therefore, the specific limitations in one or more embodiments of the anomaly diagnosis device for substation power equipment provided below can be found in the limitations of the anomaly diagnosis method for substation power equipment described above, and will not be repeated here.
[0211] In one exemplary embodiment, such as Figure 7 As shown, a fault diagnosis device for substation power equipment is provided, comprising:
[0212] The acquisition module 11 is used to acquire the operating data of the target power equipment. When abnormal features are detected in the operating data, the operating data set within a preset time period is extracted based on the abnormal time corresponding to the abnormal features. The operating data includes voiceprint data and infrared data.
[0213] Input module 12 is used to input the running dataset into the pre-built diagnostic model. The diagnostic model extracts acoustic and thermal abnormal events, matches the acoustic and thermal abnormal events based on a preset matching algorithm, obtains the quality of matched event pairs and unmatched events, and performs fusion processing based on the quality of matched event pairs and unmatched events to obtain the abnormal evolution state and diagnostic results of the target power equipment.
[0214] Each module in the aforementioned anomaly diagnosis device for substation power equipment can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.
[0215] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:
[0216] The system acquires operational data of the target power equipment. When abnormal features are detected in the operational data, the system extracts the operational dataset within a preset time period based on the abnormal time corresponding to the abnormal features. The operational data includes acoustic data and infrared data.
[0217] The running dataset is input into a pre-built diagnostic model, which extracts acoustic and thermal anomalies. Based on a preset matching algorithm, the acoustic and thermal anomalies are matched to obtain matched event pairs and unmatched event quality. The matched event pairs and unmatched event quality are then fused to obtain the abnormal evolution state and diagnostic results of the target power equipment.
[0218] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:
[0219] The system acquires operational data of the target power equipment. When abnormal features are detected in the operational data, the system extracts the operational dataset within a preset time period based on the abnormal time corresponding to the abnormal features. The operational data includes acoustic data and infrared data.
[0220] The running dataset is input into a pre-built diagnostic model, which extracts acoustic and thermal anomalies. Based on a preset matching algorithm, the acoustic and thermal anomalies are matched to obtain matched event pairs and unmatched event quality. The matched event pairs and unmatched event quality are then fused to obtain the abnormal evolution state and diagnostic results of the target power equipment.
[0221] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, performs the following steps:
[0222] The system acquires operational data of the target power equipment. When abnormal features are detected in the operational data, the system extracts the operational dataset within a preset time period based on the abnormal time corresponding to the abnormal features. The operational data includes acoustic data and infrared data.
[0223] The running dataset is input into a pre-built diagnostic model, which extracts acoustic and thermal anomalies. Based on a preset matching algorithm, the acoustic and thermal anomalies are matched to obtain matched event pairs and unmatched event quality. The matched event pairs and unmatched event quality are then fused to obtain the abnormal evolution state and diagnostic results of the target power equipment.
[0224] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.
[0225] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.
[0226] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.
Claims
1. An abnormality diagnosis method of a power device of a substation, characterized by, The method includes: The system acquires operational data of the target power equipment. If abnormal features are detected in the operational data, the system extracts an operational dataset within a preset time period based on the abnormal time corresponding to the abnormal features. The operational data includes voiceprint data and infrared data. The running dataset is input into a pre-built diagnostic model, which extracts acoustic and thermal anomalies. The acoustic and thermal anomalies are matched based on a preset matching algorithm to obtain matched event pairs and unmatched event quality. The matched event pairs and unmatched event quality are then fused to obtain the abnormal evolution state and diagnostic results of the target power equipment.
2. The method of claim 1, wherein, The diagnostic model includes a normal response model, a prediction model, a matching model, and a replacement model. The running dataset is input into the pre-constructed diagnostic model, which extracts acoustic and thermal anomalies. Based on a preset matching algorithm, the acoustic and thermal anomalies are matched to obtain matched event pairs and unmatched event quality. The matched event pairs and unmatched event quality are then fused to obtain the abnormal evolution state and diagnostic results of the target power equipment, including: The acoustic data and infrared data in the running dataset are respectively input into the normal response model, the normal response model generates acoustic residuals and infrared residuals, and the acoustic anomaly event and the thermal anomaly event are extracted from the acoustic residuals and the infrared residuals; The acoustic anomaly event is input into the prediction model to obtain the prediction result. The total mass of the acoustic event and the total mass of the thermal event are obtained by combining the prediction result, the thermal anomaly event and the predetermined hierarchical confidence level. The total mass of the acoustic events and the total mass of the thermal events are respectively input into the matching model. The matching model performs cross-modal matching according to the matching algorithm to obtain the matched event pairs, the mass of unmatched acoustic events, the mass of unmatched thermal events, and the cross-modal association confidence. The matched event pairs, the quality of the unmatched acoustic events, the quality of the unmatched thermal events, and the cross-modal correlation confidence are input into the replacement model. The replacement model replaces the acoustic residuals with the normal acoustic response under the current operating conditions and the infrared residuals with the normal infrared response under the current operating conditions. The changes in the diagnostic conclusions before and after the replacement are calculated to obtain the voiceprint evidence contribution and the infrared evidence contribution. The matched event pairs, the quality of the unmatched acoustic events, the quality of the unmatched thermal events, the cross-modal correlation confidence, the hierarchical confidence, the voiceprint evidence contribution, and the infrared evidence contribution are fused to obtain the abnormal evolution state and the diagnostic result.
3. The method of claim 2, wherein, The step of performing cross-modal matching according to the matching algorithm to obtain the matched event pairs, the quality of unmatched acoustic events, the quality of unmatched thermal events, and the cross-modal association confidence level includes: The matching cost is determined based on the time delay distribution, the characteristic distance between the acoustic anomaly and the thermal anomaly, the topological relationship between the corresponding device components of the acoustic anomaly and the thermal anomaly, and the local reliability of the acoustic anomaly and the thermal anomaly. Based on the matching algorithm and the matching cost, cross-modal matching is performed on the total quality of the acoustic events and the total quality of the thermal events to obtain the matched event pairs, the quality of the unmatched acoustic events, the quality of the unmatched thermal events, and the cross-modal association confidence.
4. The method of claim 3, wherein, The step of performing cross-modal matching on the total acoustic event quality and the total thermal event quality according to the matching algorithm and the matching cost to obtain the matched event pairs, the unmatched acoustic event quality, the unmatched thermal event quality, and the cross-modal association confidence level includes: Based on the matching algorithm, with the matching cost as the objective, the total mass of the acoustic events and the total mass of the thermal events are matched to obtain a matching matrix between the acoustic anomaly events and the thermal anomaly events; each element in the matching matrix represents the matching mass between each acoustic anomaly event and the thermal anomaly event, and acoustic anomaly events and thermal anomaly events that have not been assigned matching mass remain unmatched. According to the matching matrix, the acoustic event quality corresponding to the unmatched acoustic anomaly event is determined as the unmatched acoustic event quality, the thermal event quality corresponding to the unmatched thermal anomaly event is determined as the unmatched thermal event quality, and the acoustic anomaly event and thermal anomaly event with matching quality higher than a preset threshold are determined as the matched event pair; The reliability of the cross-modal association is determined based on the matching quality and corresponding matching cost of each event pair in all the matched event pairs.
5. The method of claim 2, wherein, The layered reliability includes acoustic modal reliability, infrared modal reliability, acoustic event local reliability, and thermal event local reliability. The method for determining the layered credibility includes: The credibility of the voiceprint modality is determined based on the quality indicators of the voiceprint data. The reliability of the infrared mode is determined based on the quality indicators of the infrared data. The local reliability of the acoustic event is determined based on the event quality index of the acoustic anomaly event; The local reliability of the thermal event is determined based on the event quality index of the thermal anomaly event.
6. The method of claim 2, wherein, The process of inputting the acoustic anomaly event into the prediction model to obtain the prediction result, and combining the prediction result, the thermal anomaly event, and a pre-determined hierarchical confidence level to obtain the total acoustic event quality and the total thermal event quality, includes: The acoustic anomaly event is input into the prediction model for prediction, and the time delay distribution and waiting probability corresponding to the acoustic anomaly event are obtained; the waiting probability is used to characterize the probability that the acoustic anomaly event will produce a corresponding thermal anomaly event after the current time. The total quality of the acoustic event is determined based on the time delay distribution, the waiting probability, and the hierarchical confidence level. The total mass of the thermal event is determined based on the infrared modal reliability of the layered reliability and the local reliability of the thermal event of the thermal anomaly.
7. The method of claim 2, wherein, The process of fusing the matched event pairs, the quality of the unmatched acoustic events, the quality of the unmatched thermal events, the cross-modal association confidence, the hierarchical confidence, the voiceprint evidence contribution, and the infrared evidence contribution to obtain the abnormal evolution state and the diagnostic result includes: Based on the matched event pairs, the matched acoustic anomaly events and thermal anomaly events are weighted and aggregated to obtain the features of the matched event pairs; The unmatched acoustic and thermal anomalies are weighted and aggregated based on the quality of the unmatched acoustic events and the quality of the unmatched thermal events, respectively, to obtain the unmatched single-modal features; The matching event pair features, the unmatched single-modal features, the hierarchical credibility, the cross-modal association credibility, the voiceprint evidence contribution, and the infrared evidence contribution are concatenated to obtain the fused features; The fusion features are subjected to state discrimination processing to obtain the abnormal evolution state and the diagnostic result.
8. An abnormality diagnosing device of a power equipment of a substation, characterized by comprising: The device includes: The acquisition module is used to acquire the operating data of the target power equipment. When an abnormal feature is detected in the operating data, the operating data set within a preset time period is extracted based on the abnormal time corresponding to the abnormal feature. The operating data includes voiceprint data and infrared data. The input module is used to input the running dataset into a pre-built diagnostic model, which extracts acoustic and thermal anomalies, matches the acoustic and thermal anomalies based on a preset matching algorithm, obtains matched event pairs and unmatched event quality, and performs fusion processing based on the matched event pairs and unmatched event quality to obtain the abnormal evolution state and diagnostic results of the target power equipment. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.
10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.
