A fault detection method based on multi-source time sequence signals, electronic equipment, computer readable storage medium and program product

Abstract

This application relates to the field of fault detection technology, providing a fault detection method, electronic device, computer-readable storage medium, and program product based on multi-source time-series signals. The method can be applied to fault detection of elevator drum brakes, including: acquiring multi-source time-series signals of the elevator drum brake within a single operating cycle; constructing a graph structure based on the multi-source time-series signals; obtaining an energy difference coefficient based on the multi-source time-series signals and the graph structure, whereby the energy difference coefficient characterizes the degree of energy difference between electrical energy consumption and mechanical energy generation between adjacent operating events; inputting the graph structure and the energy difference coefficient into a deep learning model to obtain a graph embedding vector; and obtaining a fault probability value based on the graph embedding vector, whereby the fault probability value is used to determine the fault status of the elevator drum brake. This application can improve the monitoring sensitivity and detection accuracy of early local faults in elevator drum brakes.

Description

Technical Field

[0001] This invention relates to the field of fault detection technology, and more specifically, to a fault detection method, electronic device, computer-readable storage medium, and program product based on multi-source timing signals. Background Technology

[0002] In industrial production, transportation, equipment manufacturing, and automation control applications, electromechanical equipment is frequently used. This equipment typically includes electrically driven components, mechanical transmission components, and actuators. During operation, the electrically driven components receive electrical energy and generate a mechanical response through transmission structures and actuators. Over long-term operation, local components within the equipment may experience early localized failures due to jamming, wear, or other reasons. In their initial stages, these failures usually only affect localized mechanical processes, such as delayed local action response, insufficient mechanical displacement, or abnormal electrical drive input. At this point, the equipment as a whole can still complete its operation. However, as the equipment's lifespan increases, these minor localized failures can develop into overall equipment malfunctions. More importantly, these early, minor localized failures are difficult to identify directly using the characteristic values ​​of a single or multiple signals.

[0003] In existing technologies, time-series analysis is often performed based on one or more electrical and mechanical signals, and coefficients or thresholds are used to determine whether equipment operation is abnormal. While these methods can reflect some mechanical anomalies, they typically focus on continuously sampled signals, individual motion characteristics, or overall operating parameters, thus only reflecting faults in the overall operating state. Therefore, existing technologies have poor ability to detect minor and localized anomalies in the early stages of fault development during equipment operation, easily leading to missed detections. Therefore, there is an urgent need for a fault detection method that can improve the sensitivity of detecting early localized faults in equipment, thereby improving the accuracy of fault detection.

[0004] Taking an elevator drum brake as an example, an elevator drum brake typically involves multiple continuous stages, including the energization and de-energization of the electromagnetic coil, current changes, the movement of the push rod and brake arm, brake release and engagement, and the formation of braking force. When impurities are present in the electromagnet inside the brake, oxidation, corrosion, and sludge buildup can cause jamming. This prevents the push rod from being pushed back to its original position when the elevator stops, resulting in poor or no contact between the brake lining and the brake wheel, causing a decrease or even loss of the elevator's braking force, and potentially leading to a safety accident. Summary of the Invention

[0005] In view of this, the purpose of the embodiments of this application is to provide a fault detection method, electronic device, computer-readable storage medium and program product based on multi-source timing signals, which can improve the monitoring sensitivity of early local faults of the device under test and improve the detection accuracy.

[0006] To achieve the above technical objectives, the technical solution adopted in this application is as follows:

[0007] In a first aspect, embodiments of this application provide a fault detection method based on multi-source time-series signals, the method comprising:

[0008] Acquire multi-source timing signals of the device under test within a single operating cycle. The multi-source timing signals include voltage signals, current signals, displacement signals, and force signals during the operation of the device under test.

[0009] Based on the multi-source time-series signals, a graph structure is constructed. The nodes of the graph structure are the working events within the operating cycle of the device under test. The edges of the graph structure represent the state change process between two working events. The node features include the occurrence time of the working event and the signal amplitude corresponding to the occurrence time. The signal amplitude includes at least one of the amplitude of a voltage signal, the amplitude of a current signal, the amplitude of a displacement signal, and the amplitude of a force signal. The edge features include a first difference between the occurrence times of two working events and a second difference between the signal amplitudes at the occurrence times of two working events.

[0010] Based on the graph structure, an energy difference coefficient is obtained by measuring the changes in electrical energy and mechanical energy between the two nodes. The energy difference coefficient is used to characterize the degree of energy difference between the electrical energy consumed and the mechanical energy generated by the device under test between the two working events. Both the changes in electrical energy and the changes in mechanical energy are obtained based on the node characteristics of the graph structure.

[0011] The graph structure and the energy difference coefficient are input into a deep learning model. Based on the graph structure and the energy difference coefficient, the deep learning model obtains a fault probability value, which is used to determine the fault status of the device under test.

[0012] According to the first aspect, the energy difference coefficient, obtained based on the graph structure and through the changes in electrical energy and mechanical energy between the two nodes, includes:

[0013] The change in electrical energy is obtained based on the amplitude of the voltage signal and the amplitude of the current signal in the node characteristics.

[0014] The change in mechanical energy is obtained based on the amplitude of the force signal and the amplitude of the displacement signal in the node features.

[0015] Based on the changes in electrical energy and mechanical energy, an energy difference coefficient is obtained according to the energy conversion rate of the device under test. The energy conversion rate is obtained based on the multi-source time-series signal.

[0016] According to the first aspect, the deep learning model includes a graph Transformer network (a deep learning model architecture that extends the self-attention mechanism of the Transformer architecture to graph-structured data in non-Euclidean space), the graph Transformer network including a graph attention layer, a readout layer, and a classification output layer, and the deep learning model obtains a fault probability value based on the graph structure and the energy difference coefficient, including:

[0017] The graph structure and the energy difference coefficient are input into the graph attention layer, and the deep node features are obtained through the graph attention layer.

[0018] The deep node features are input to the readout layer, and the readout layer concatenates the deep node features to obtain the graph embedding vector;

[0019] The graph embedding vector is input into the classification output layer, and the fault probability value is obtained through the classification output layer. The fault probability value is used to determine the fault detection result.

[0020] According to the first aspect, obtaining deep node features through the graph attention layer includes:

[0021] Based on the edge features, each feature in the edge features is enhanced using the energy difference coefficient to obtain enhanced edge features. The edge features include at least one feature, which is obtained by the first difference and the second difference.

[0022] The node features and the enhanced edge features are input into the graph attention layer, and the graph attention layer is used to obtain the first attention score of the node to itself and the second attention score of the node to its neighboring nodes.

[0023] Normalize the first attention score and the second attention score to obtain the first attention weight of the node to itself and the second attention weight of the node to its neighboring nodes;

[0024] Based on the first attention weight and the second attention weight, the deep node features of the node are obtained through weighted aggregation.

[0025] According to the first aspect, the energy conversion efficiency is obtained based on the multi-source time-series signal, including:

[0026] Based on the voltage signal and the current signal in the multi-source time-series signal, the electrical energy consumed by the device under test in a single operating cycle is obtained by integrating the voltage signal and the current signal over a single operating cycle.

[0027] Based on the force signal and displacement signal in the multi-source time-series signal, the mechanical energy generated by the device under test in a single operating cycle is obtained by integrating the force signal and displacement signal over time in a single operating cycle.

[0028] The energy conversion rate is obtained based on the electrical energy and the mechanical energy.

[0029] According to the first aspect, prior to acquiring the multi-source timing signals of the device under test within a single operating cycle, the method further includes:

[0030] Acquire multiple historical multi-source time-series signals and the corresponding historical real fault results of the historical multi-source time-series signals. The historical multi-source time-series signals include historical voltage signals, historical current signals, historical displacement signals, and historical force signals of the equipment during historical operation.

[0031] Based on the historical multi-source time-series signals, a historical graph structure is constructed. The nodes of the historical graph structure are the working events within the operating cycle of the device under test. The edges of the historical graph structure represent the state change process between two working events. The historical node features of the historical graph structure include the historical occurrence time of the working event and the historical signal amplitude corresponding to the historical occurrence time. The historical signal amplitude includes at least one of the amplitude of historical voltage signal, historical current signal, historical displacement signal, and historical force signal. The historical edge features of the historical graph structure include the historical first difference corresponding to the historical occurrence time of two working events and the historical second difference between the historical signal amplitudes of the two working events at the historical occurrence time.

[0032] Based on the historical graph structure, the historical energy difference coefficient is obtained by using the historical changes in electrical energy and mechanical energy between the two nodes.

[0033] The historical graph structure, the historical energy difference coefficient, and the historical real fault results are integrated into training data.

[0034] The deep learning model is trained and initialized based on the training data, so that after the graph structure and the energy difference coefficient are input into the deep learning model, the deep learning model outputs the fault probability value.

[0035] Secondly, embodiments of this application also provide an electronic device, the electronic device including a processor and a memory coupled to each other, the memory storing a computer program, and when the computer program is executed by the processor, causing the electronic device to perform the method as described in the first aspect.

[0036] Thirdly, embodiments of this application also provide a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, which, when run on a computer, causes the computer to perform the method described in the first aspect.

[0037] Fourthly, embodiments of this application also provide a program product, characterized in that it includes a computer program, which, when executed by a processor, implements the method described in the first aspect.

[0038] The invention employing the above technical solution has the following advantages:

[0039] In the technical solution provided in this application, by constructing a graph structure with working events as nodes and the state change processes between adjacent working events as edges, the overall operation process of the device under test is characterized as a continuous action process with state change relationships. This is beneficial for identifying fault accumulation phenomena such as insufficient internal displacement and slow mechanical force response, and for extracting fault features that gradually accumulate under different operating states. By constructing an energy difference coefficient, the model can focus on the degree of difference between electrical energy consumption and mechanical energy generation between adjacent working events. When making fault judgments, the model not only focuses on changes in various signals, but also on whether the electric drive input and the mechanical execution result are coordinated, which helps to improve the ability to identify weak faults and improve the accuracy of fault detection. Attached Figure Description

[0040] This application can be further illustrated by the non-limiting embodiments given in the accompanying drawings. It should be understood that the following drawings only illustrate some embodiments of this application and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained from these drawings without any inventive effort.

[0041] Figure 1 A flowchart of a fault detection method based on multi-source timing signals provided in an embodiment of this application. Detailed Implementation

[0042] The present application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that similar or identical parts are referred to by the same reference numerals in the drawings or description. Implementations not shown or described in the drawings are forms known to those skilled in the art. In the description of this application, terms such as "first" and "second" are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0043] The device under test or equipment described in this embodiment takes an elevator drum brake as an example.

[0044] Please refer to Figure 1This application provides a fault detection method based on multi-source timing signals, which can be applied to electronic devices and whose steps can be executed or implemented by the electronic device. The electronic device can be, but is not limited to, personal computers, smartphones, and other electronic devices. The fault detection method based on multi-source timing signals may include the following steps:

[0045] S10 acquires multi-source timing signals of the elevator drum brake during a single operating cycle. The multi-source timing signals include voltage signals, current signals, displacement signals, and force signals during the operation of the elevator drum brake.

[0046] S20 constructs a graph structure based on the multi-source time-series signal. The nodes of the graph structure are the working events within the operating cycle of the elevator drum brake. The edges of the graph structure represent the state change process between two working events. The node features include the occurrence time of the working event and the signal amplitude corresponding to the occurrence time. The signal amplitude includes at least one of the amplitude of a voltage signal, the amplitude of a current signal, the amplitude of a displacement signal, and the amplitude of a force signal. The edge features include a first difference between the occurrence times of two working events and a second difference between the signal amplitudes of the occurrence times of two working events.

[0047] Based on the graph structure, S30 obtains an energy difference coefficient through the changes in electrical energy and mechanical energy between the two nodes. The energy difference coefficient is used to characterize the degree of energy difference between the electrical energy consumed and the mechanical energy generated by the elevator drum brake between the two working events. Both the changes in electrical energy and the changes in mechanical energy are obtained based on the node characteristics of the graph structure.

[0048] S40 inputs the graph structure and the energy difference coefficient into the deep learning model. The deep learning model obtains a fault probability value based on the graph structure and the energy difference coefficient. The fault probability value is used to determine the fault status of the elevator drum brake.

[0049] This application acquires multi-source time-series signals, including voltage, current, displacement, and force signals, within a single operating cycle of an elevator drum brake. Based on these multi-source time-series signals, a graph structure is constructed. Further, based on the multi-source time-series signals and the graph structure, the degree of energy difference between adjacent operating events (electrical energy consumption and mechanical energy generation) is determined, yielding an energy difference coefficient. Finally, the graph structure and energy difference coefficient are input into a deep learning model, enabling the model to learn the characteristics of different operating events, the state change process characteristics between two operating events, and the impact of energy differences on the process. Ultimately, a fault probability value is obtained, and the fault status of the elevator drum brake is determined based on this probability value. This application improves the accuracy of fault detection for elevator drum brakes by establishing a graph structure and energy difference coefficient, and by employing a deep learning model to learn the influence of energy differences on the operation process of the elevator drum brake.

[0050] The steps of the fault detection method based on multi-source time-series signals will be described in detail below:

[0051] In S10, when acquiring the multi-source timing signals of the elevator drum brake, the methods include, but are not limited to, the following: A voltage sensor is connected in parallel to the power supply terminal of the electromagnet coil to collect the voltage across the coil in real time, obtaining a voltage signal U(t). Similarly, a Hall current sensor is installed on the power supply line of the electromagnet coil to collect the coil current in real time, obtaining a current signal I(t). A photoelectric displacement sensor or linear displacement sensor is installed in the direction of movement of the push rod, armature, or brake arm to collect the opening and closing displacements of the push rod in real time, obtaining a displacement signal S(t). A force sensor is installed on the force transmission path of the push rod and brake arm to collect the force experienced by the push rod during opening and closing, obtaining a force signal F(t). Each sensor is connected to a data acquisition module, which synchronously acquires the voltage signal, current signal, displacement signal, and force signal at a unified sampling frequency, and forms a multi-source timing signal within a single operating cycle based on a unified timestamp. A single operating cycle can be defined as the time required for the electromagnet coil to go from being energized to de-energized and complete the push rod return to its original position in a single operation. Specifically, the start time of the single operating cycle can be determined when the coil voltage rises to a preset start threshold, and the end time of the single operating cycle can be determined when the coil voltage drops to a preset stop threshold and the displacement signal stabilizes at the closed position.

[0052] In S20, a graph structure is constructed based on the multi-source time-series signal, including:

[0053] S21 determines multiple working events of the elevator drum brake based on different states in the preset operation process of the elevator drum brake, and uses the working events as the nodes;

[0054] S22 obtains the edge based on the temporal relationship between two adjacent work events. The edge is a directed edge, with the direction pointing from the node corresponding to the earlier work event to the node corresponding to the later work event, and is used to represent the work process between adjacent work events.

[0055] S23 concatenates the occurrence time corresponding to the occurrence of the working event and the signal amplitude corresponding to the occurrence time to obtain the node feature.

[0056] S24 concatenates the first difference between the occurrence times of two adjacent working events with the second difference between the signal amplitudes at the occurrence times of two adjacent working events to form the edge feature;

[0057] S25 integrates the nodes, edges, node features, and edge features into the graph structure.

[0058] In S21, the selection of working events is based on the preset operating flow of the equipment, and multiple key state points are predefined. In this embodiment, the working principle of the elevator drum brake is as follows: After the elevator starts, the internal coil of the electromagnet is energized, allowing current to flow through it. This induces a magnetic field in the internal armature, causing them to repel each other and push the push rod outward, thus disengaging the brake arm assembly and brake liner from the brake wheel. At this time, there is no braking force. When the elevator stops, the electromagnet is de-energized, and under the action of the compression spring, the push rod and brake arm assembly are pushed back to their original positions. The brake liner contacts the brake wheel, generating braking force to stop the elevator.

[0059] The working events described in this embodiment specifically include the following six events: coil energization event (D1), current peak event (D2), push rod start event (D3), push rod arrival event (D4), gate closing start event (D5), and gate closing completion event (D6). The coil energization event (D1) represents the response of the electromagnet coil starting to be energized, extracted through the coil current signal. The occurrence time of this event is determined when the coil current rises from zero and exceeds a preset energization current threshold. The current peak event (D2) represents the response of the coil current reaching its maximum value during the gate opening phase, extracted through extreme value detection of the coil current signal. The occurrence time of this event is determined when the current rises to its highest point during the gate opening phase and begins to fall. This event corresponds to the critical state where the electromagnetic attraction force rises rapidly and is about to drive mechanical action. The push rod start event (D3) represents the initial response of the mechanical action during the gate opening phase, extracted through the push rod displacement signal. The occurrence time of this event is determined when the push rod displacement changes from its initial zero position and exceeds a preset displacement start threshold. The event corresponds to the state transition point where the push rod begins to move. The push rod reaching its position event (D4) represents the response to the completion of the opening action. It is extracted through the push rod displacement signal. When the push rod displacement reaches its maximum stroke and enters a stable fluctuation range with a change less than the preset reaching threshold, this event is determined to have occurred. This event corresponds to the starting point of the elevator drum brake fully opening and the elevator being able to operate normally. The closing start event (D5) represents the initial response to the closing action. It is extracted through the displacement signal. When the push rod displacement begins to retract from its maximum stroke to the initial position, this event is determined to have occurred. This event corresponds to the state transition point where the coil loses power, the electromagnetic attraction weakens, and the spring force begins to drive the push rod to reset. The closing completion event (D6) represents the response to the completion of the closing action and the establishment of braking force. It is extracted through the combined push rod displacement signal and braking force signal. When the push rod displacement returns to the initial zero position and the braking force rises to a stable rated value range with a fluctuation less than the preset stable threshold, this event is determined to have occurred. This event corresponds to the final state where the elevator drum brake is fully closed and the braking force is fully established. The node mapping method of the graph structure is to correspond each working event to an independent node in the graph structure. The timing numbers of the six nodes (D1-D6) correspond to the order of actions in the braking process, and each node carries the equipment operating status information at the corresponding moment.

[0060] In S22, adjacency relationships are established for each node according to the chronological order of the work events. The first and last nodes (D1, D6) have only one adjacent node in their subsequent and preceding orders, respectively, while the middle nodes (D2~D5) each have two adjacent nodes in their preceding and following orders, forming a chain-like graph topology, corresponding to the unidirectional temporal progression logic of the braking process. That is, the following directed edges are established:

[0061]

[0062] In this diagram, the direction of each edge points from the node corresponding to the earlier working event to the node corresponding to the later working event. The five edges correspond to the following working processes: the electromagnetic response process from when the coil is first energized to when the current is established (D1-D2); the process of pushing the push rod to start moving after the electromagnetic response is established (D2-D3); the mechanical movement process of the push rod moving from the start to the open position (D3-D4); the holding process from when the gate is in the open state to when the coil is de-energized (D4-D5); and the process of the push rod returning to the closed position and the gate closing after the coil is de-energized (D5-D6). If a jamming fault exists in D3-D4, it will often manifest as: a longer duration of this process, insufficient displacement change, abnormally increased force signal, and high electrical input but insufficient mechanical displacement output, among other abnormal phenomena.

[0063] In S23, for any work event The time of its occurrence is The amplitudes of voltage, current, displacement, and force signals are read from the multi-source timing signals after timing synchronization as the signal amplitudes at the time of occurrence. That is, the signal amplitudes include the amplitudes of the voltage signal, current signal, displacement signal, and force signal. The node features are obtained by splicing. .

[0064] In S24, edge features represent the response process between two adjacent working events. Each edge corresponds to a 5-dimensional feature vector, with dimensions corresponding to the node features, representing the amount of change between adjacent nodes. Specifically, This is a current work-related event. This is the second work event. Calculate the time difference between the occurrence times of the two events as the first difference value. Calculate the difference in signal amplitudes corresponding to the two events: , , , The difference in signal amplitude is used as the second difference. The first difference and the second difference are concatenated to form an edge feature: Assuming there is a jamming fault in D3-D4, then The larger the size, the slower the opening process. If it becomes smaller, it means the displacement is not complete; An increase indicates abnormal resistance; An anomaly indicates that the electromagnet's current response is unstable.

[0065] In S25, for this embodiment, the diagram structure can be represented as follows:

[0066]

[0067] Where G represents the graph structure, D represents the set of nodes, E represents the set of edges, H represents the set of node features, and R represents the set of edge features. , , , .

[0068] Existing technologies typically focus only on the signal amplitude at a single point in time, or only on the overall temporal characteristics, neglecting the continuous state changes during the operation of an elevator drum brake. Taking the elevator drum brake of this application as an example, when an early, minor fault occurs, the current, voltage, displacement, or force at a certain moment may all be within the normal range. However, if, after the coil is energized, the push rod starts moving too slowly, the opening displacement is insufficient, or the required force increases abnormally, it may indicate that there is internal obstruction in the electromagnet. This abnormality is not necessarily reflected in the absolute value of a single signal, but is more likely to be reflected in the time changes, displacement changes, and force changes between different operating events. This application decomposes the operation process of the elevator drum brake into a combination of discrete states and continuous processes, reducing data dimensionality and computational complexity. It also facilitates subsequent deep learning models in learning the inherent laws governing the interaction between electrical and mechanical energy during each event, thereby improving the accuracy of detecting early, minor faults.

[0069] In S30, based on the graph structure, the energy difference coefficient is obtained through the changes in electrical energy and mechanical energy between the two nodes, including:

[0070] S31 obtains the change in electrical energy based on the amplitude of the voltage signal and the amplitude of the current signal in the node characteristics;

[0071] S32 obtains the change in mechanical energy based on the amplitude of the force signal and the amplitude of the displacement signal in the node features;

[0072] S33 obtains an energy difference coefficient based on the change in electrical energy and the change in mechanical energy, according to the energy conversion rate of the elevator drum brake. The energy conversion rate is obtained based on the multi-source time-series signal.

[0073] In S31, for nodes i and j, the electrical power of the two nodes is calculated using the trapezoidal integral approximation to obtain the corresponding change in electrical energy:

[0074]

[0075]

[0076]

[0077] in, Indicates work event When it occurs, the electrical power of the electromagnet coil, Representing work events The amplitudes of the voltage and current signals corresponding to the moment of occurrence; Indicates work event When it occurs, the electrical power of the electromagnet coil, Representing work events The amplitude of the voltage signal and the amplitude of the current signal corresponding to the moment of occurrence. This represents the change in electrical energy corresponding to node i and node j. This represents the time difference between the times at which two events occur. (The aforementioned...) It is a simplified way of realizing the change in electrical energy, and its basic physical basis is the product of electrical power and time.

[0078] In another implementation, for operating conditions with high sampling frequencies, numerical integration can also be performed using continuously acquired voltage and current signals:

[0079]

[0080] in, Representing an edge The corresponding change in electrical energy represents the change in voltage and current over time. to Integral over an interval. Indicates voltage signal, Represents a current signal. and Let i and j represent the occurrence times of the i-th node and the j-th node, respectively.

[0081] In S32, for nodes i and j, calculate the change in mechanical energy between the two nodes.

[0082]

[0083]

[0084] in, This represents the change in mechanical energy corresponding to node i and node j. Representing work events and The force at the time of occurrence Representing work events and Displacement at the time of occurrence Representing an edge The absolute value of the displacement change. This is a simplified way of realizing the change in mechanical energy. Without considering potential energy, the basic physical basis is the work done by the force along the displacement direction. In another implementation, for operating conditions with higher sampling frequencies, continuous signal integration can also be used:

[0085]

[0086] in This represents the change in mechanical energy corresponding to node i and node j. Indicates force signal, Indicates displacement signal, and Let i and j represent the occurrence times of the i-th node and the j-th node, respectively.

[0087] In S33, before calculating the energy difference coefficient, the edge formed by node i and node j can also be calculated. The proportion of electrical energy change and the proportion of mechanical energy change.

[0088]

[0089]

[0090] in, Representing an edge The percentage of change in electrical energy represents the current side within a single operating cycle. The proportion of all changes in side energy. Representing an edge The corresponding change in electrical energy. Representing an edge The percentage of mechanical energy change represents the current side within a single operating cycle. The proportion of all changes in side mechanical energy. Representing an edge The corresponding change in mechanical energy.

[0091] Under normal circumstances, if a certain side consumes a lot of electrical input, it should also generate a corresponding mechanical action output.

[0092] Based on the proportions of electrical energy change and mechanical energy change, and according to the energy conversion rate of the elevator drum brake, the energy difference coefficient is calculated.

[0093]

[0094] The energy difference coefficient is a coefficient constructed based on the coordinated relationship between changes in electrical energy and mechanical energy, and is calculated based on multi-source time-series signals within the current period. For the edge... Energy difference coefficient Calculate the edges Percentage of changes in electrical energy With edge proportion of mechanical energy change The difference between the first and second differences is taken as the first difference. Then, the difference between unit 1 and the energy conversion rate is calculated and taken as the second difference. The product of the first and second differences is used as a gating term, and the sum of the gating term and unit 1 is taken as the energy difference coefficient. Based on the energy difference coefficient of each edge, a set of energy difference coefficients is obtained. .

[0095] The proportions of electrical energy change and mechanical energy change represent the electrical input contribution and mechanical output contribution of the edge throughout the entire action cycle. The product of the first and second differences is used as a gating term, a joint gating term that simultaneously satisfies two conditions: "an overall trend of energy loss" and "a local mismatch between electrical and mechanical energy at the current edge." When the overall energy conversion is basically normal, even if there are slight inconsistencies in the local area due to sampling fluctuations, the energy difference coefficient will not significantly increase; when there is overall energy loss but no obvious mechanical output deficiency in the current interval, the overall anomaly will not be indiscriminately applied to that interval. Only when both conditions exist simultaneously will the influence of the corresponding edge features in the attention calculation be increased. The energy difference coefficient serves to provide dimensionless enhancement to the edge features, rather than representing physical energy or fault determination results.

[0096] The energy difference coefficient reflects the overall match between electrical energy consumption corresponding to voltage and current and mechanical energy generation corresponding to displacement and force, indicating the degree of difference between electric drive and mechanical action. In practical applications, during the early stages of faults or abnormal conditions, elevator drum brakes often do not individually exceed threshold values ​​for current peaks, voltage amplitudes, displacement values, or force values. For example, the coil current may establish normally, but the brake lever may fail to move in time due to a jamming fault, or although the brake lever may produce a small displacement, it may not generate mechanical energy matching the electrical energy consumption. Setting thresholds only for individual signal characteristics can easily miss early fault situations where the electrical signal appears normal but the actual mechanical action is insufficient. The energy difference coefficient maps electrical signals and mechanical signals to the correspondence between electrical energy consumption and mechanical energy generation, enabling the identification of mismatches where electrical input has occurred but mechanical output is insufficient. In the event of a jamming fault, even if the coil has consumed a significant amount of electrical energy, the mechanical movement may be obstructed, preventing the corresponding displacement or force output, thus resulting in an increased energy difference coefficient. Therefore, the energy difference coefficient makes edges with high power consumption but insufficient mechanical output receive more attention in subsequent graph attention calculations, thereby improving the model's accuracy in identifying jamming faults and reducing the possibility of misjudgment caused by simple electrical signal fluctuations or instantaneous mechanical disturbances.

[0097] In S33, the energy conversion rate is obtained based on the multi-source time-series signals, including: based on the voltage signal and the current signal in the multi-source time-series signals, by integrating the voltage signal and the current signal over time within a single operating cycle, the electrical energy consumed by the elevator drum brake in a single operating cycle is obtained; based on the force signal and the displacement signal in the multi-source time-series signals, by integrating the force signal and the displacement signal over time within a single operating cycle, the mechanical energy generated by the elevator drum brake in a single operating cycle is obtained; based on the electrical energy and the mechanical energy, the energy conversion rate is obtained. Since event D1 is the first event and D6 is the last event, D1 and D6 are used as the start and end nodes of a single operating cycle, respectively.

[0098]

[0099]

[0100]

[0101] Wherein, En1 represents mechanical energy and En2 represents electrical energy; The energy conversion rate characterizes the proportion of input electrical energy converted into effective mechanical energy output. t1 and t6 represent the occurrence times of event D1 and event D6, respectively. This represents the force signal between events D1 and D6. This represents the displacement signal between events D1 and D6. These represent the voltage and current signals between events D1 and D6, respectively.

[0102] In S40, the deep learning model includes a graph Transformer network, which comprises a graph attention layer, a readout layer, and a classification output layer. It should be noted that the graph attention layer used in this embodiment is a single-head graph attention layer. The deep learning model obtains the fault probability value based on the graph structure and the energy difference coefficient, including:

[0103] S41 inputs the graph structure and the energy difference coefficient into the graph attention layer, and obtains the deep node features through the graph attention layer;

[0104] S42 inputs the deep node features to the readout layer, and the readout layer concatenates the deep node features to obtain the graph embedding vector;

[0105] S43 inputs the graph embedding vector to the classification output layer, and obtains the fault probability value through the classification output layer. The fault probability value is used to determine the fault detection result.

[0106] In S41, deep node features are obtained through the graph attention layer, including:

[0107] S411 Based on the edge features, through the energy difference coefficient, feature enhancement is performed on each feature in the edge features to obtain enhanced edge features. The edge features include at least one feature, which is obtained by the first difference and the second difference.

[0108] S412 inputs the node features and the enhanced edge features into the graph attention layer, and obtains the first attention score of the node to itself and the second attention score of the node to its neighboring nodes through the graph attention layer;

[0109] S413 normalizes the first attention score and the second attention score to obtain the first attention weight of the node to itself and the second attention weight of the node to its neighboring nodes;

[0110] S414 obtains the deep node features of the node through weighted aggregation based on the first attention weight and the second attention weight.

[0111] In S411, the edge features of the graph structure can be represented as: Specifically, for the edge edge features It can be represented as The energy difference coefficient is used to enhance the features of each dimension in the edge features.

[0112]

[0113]

[0114]

[0115]

[0116]

[0117]

[0118]

[0119] in, Indicates enhanced edge features, Indicates the energy difference coefficient. These represent the first and second differences in the preceding features, namely: the difference at the time of occurrence, the difference in voltage signal amplitude, the difference in current signal amplitude, the difference in displacement signal amplitude, and the difference in force signal amplitude. These represent the differences in the timing of the enhancement, the amplitude differences in the enhanced voltage signal, the amplitude differences in the enhanced current signal, the amplitude differences in the enhanced displacement signal, and the amplitude differences in the enhanced force signal, respectively, within the enhanced edge features.

[0120] Edge features are enhanced by using the energy difference coefficient corresponding to the same side, thus amplifying the feature amplitude of edges with significant energy differences. This enhancement preserves the relationships between work events in terms of time, electrical energy, and mechanical energy, making edges with large energy differences have a stronger impact in subsequent attention score calculations. This makes the model more likely to focus on intervals where there may be motion lag, insufficient displacement, or abnormal force. Consequently, the accuracy of identifying gradual and localized faults is improved.

[0121] In S412, the candidate nodes for which attention needs to be calculated for the current node are: the current node itself and its neighboring nodes. The neighboring nodes include the previous neighboring node and the next neighboring node that are directly connected to the current node through an edge.

[0122] For the current node and candidate nodes It is necessary to determine the enhancement edge features used by the two in the attention calculation. In order to make the meaning clear when calculating graph attention, the enhancement features can be uniformly represented as:

[0123]

[0124] The node characteristics corresponding to the i-th node are: . This indicates the time when the i-th work event occurs; This indicates the voltage signal amplitude corresponding to the moment of occurrence; This indicates the amplitude of the current signal at the moment of occurrence; This indicates the amplitude of the displacement signal corresponding to the moment of occurrence; This indicates the amplitude of the force signal at the moment of occurrence.

[0125] Calculate attention score:

[0126]

[0127]

[0128]

[0129]

[0130] in, Let represent the query vector for the i-th node. Let represent the key vector of the i-th node. This represents the value vector of the i-th node. , , These represent the query mapping matrix, key mapping matrix, and value mapping matrix, respectively. , , These represent the corresponding bias vectors; , , , , , These are all trainable parameters of the model. Indicates the current node For candidate nodes Attention score; Indicates candidate nodes The key vector; Indicates the dimensions of the query vector and key vector; This represents the edge feature weight vector, which consists of parameters obtained during model training. This represents the enhanced edge features between the current node and the candidate nodes.

[0131] Specifically, when calculating the first attention score, since the candidate node for attention calculation at this time is the node itself, Therefore, for the current node i, the first attention score to itself is... , When calculating the second attention score, since the candidate nodes for attention calculation are adjacent nodes, at this time... Therefore, for the second attention score of the current node i to its neighboring nodes, , .

[0132] During the attention score calculation process It is used to calculate the correlation between the voltage, current, displacement, and force of the current node and the candidate nodes at the moment of occurrence. To enhance the impact of edges on the attention score, a term is included that reflects the changes in time, voltage, current, displacement, and force between the current node and its neighboring nodes after being amplified by energy differences. Therefore, energy differences can influence the current node's level of attention to its neighboring nodes. Furthermore, when the candidate node is the current node itself... Take the zero vector and calculate the attention score based solely on the state association between the current node and its own node.

[0133] In S413, the first attention weight and the second attention weight are calculated, and both the first attention score and the second attention score are subjected to Softmax normalization (a normalization exponential function that compresses a K-dimensional real vector into another K-dimensional real vector, such that the value of each element is in the interval (0,1) and the sum of all elements is 1):

[0134]

[0135]

[0136]

[0137] in, Indicates the current node The weight of one's own primary attention. This represents the first attention score. Indicates the current node Second attention weights for neighboring nodes. Represents the set of adjacent nodes. This indicates the second attention score. This represents an exponential function.

[0138] In S414, based on the first attention weight and the second attention weight, the deep node features of the node are obtained through weighted aggregation.

[0139]

[0140] in, Indicates the current node The deep node features obtained after graph attention processing; Represents the set of adjacent nodes; Indicates the current node The weight of one's own primary attention; Indicates the current node Second attention weights for neighboring nodes; Indicates adjacent nodes The value vector is obtained by value mapping from the node features of the candidate node; Represents the value vector of the current node; This represents a non-linear activation function, which can be the ReLU activation function (Rectified Linear Unit, one of the activation functions currently used in deep neural networks).

[0141] The larger the attention weight of a candidate node, the more its value vector relates to the features of deep nodes. The greater the contribution, the smaller the attention weight of a candidate node, and the more effective its corresponding state information becomes. The weaker the influence, the better. In this application, the second attention weight is affected by the enhanced edge features. The enhanced edge features include edge features enhanced by the energy difference coefficient. Therefore, when there is an energy difference between adjacent working events where the electrical input contribution is high but the mechanical output is insufficient, the corresponding edge features participate more in the calculation of the second attention score and further affect the second attention weight. Finally, the deep node features of the current node not only retain the state information of the current working event itself, but also integrate the information of the change process of adjacent working events related to the fault.

[0142] The following uses nodes For example, its candidate nodes include the current node itself and its neighboring nodes:

[0143]

[0144] in, This represents the set of candidate nodes corresponding to the current node 4, consisting of its own node and neighboring nodes. The preceding adjacent node, For the current node itself, It is the next adjacent node in the subsequent order. That is, When calculating attention weights, it is necessary to calculate their effect on themselves separately. The first attention score, and the , The second attention score is then normalized.

[0145] First, obtain the node features corresponding to the three candidate nodes:

[0146]

[0147]

[0148]

[0149] in, Indicates the time when the work event occurred. , , , These represent the voltage, current, displacement, and force signal amplitudes at that moment, respectively. The characteristics of each node reflect the instantaneous operating state when the corresponding working event occurs. Then, using the query mapping matrix, key mapping matrix, and value mapping matrix in the graph attention layer, the query vector, key vector, and value vector of the three nodes are calculated respectively:

[0150]

[0151]

[0152]

[0153] in, ; For nodes The query vector represents the information that the node needs to pay attention to when updating its own features; The key vector represents the state information of a node that can be matched by other nodes; It is a value vector representing the information that can be weighted and aggregated for this node in the future; , , and , , All of these are model parameters obtained during the training process.

[0154] Next, based on the enhanced edge features corresponding to adjacent nodes. For edges... Its original edge characteristics are:

[0155]

[0156] Its corresponding enhanced edge features are: in, Indicates work event The energy difference coefficient, This represents the enhanced edge feature after being strengthened by the energy difference coefficient.

[0157] Similarly, regarding work-related events :

[0158]

[0159] Indicates from arrive Strengthened edge features, Indicates from arrive The enhanced edge features.

[0160] For the current node itself The first attention score is:

[0161]

[0162] For the preceding node With subsequent nodes The second attention score is:

[0163]

[0164] After obtaining the first attention score and the second attention score, perform Softmax normalization on them:

[0165]

[0166]

[0167]

[0168] Finally, the value vectors of the three candidate nodes are weighted and aggregated using three attention weights:

[0169]

[0170] In S42, the deep node features corresponding to each working event are concatenated sequentially to obtain the graph embedding vector:

[0171]

[0172] in, The graph embedding vector is used to characterize the overall operating state of the elevator drum brake within a single operating cycle. This represents the concatenation function (which connects the deep node features of each node in a preset order to form a fixed-dimensional vector that can represent global information). These represent the features of each depth node arranged according to a preset sequence of working events; This represents the number of nodes in the graph structure, i.e., the number of worker events. For this embodiment, each depth node feature... All dimensions are Then the graph embedding vector The dimension is .

[0173] For minor anomalies in the early stages of fault development, the push rod typically only exhibits slight lag or insufficient displacement within a localized action range. At this time, individual indicators such as peak current, maximum displacement, and maximum force may still be within normal fluctuation ranges, and the overall braking behavior can still be completed. Therefore, traditional threshold judgment is prone to missing detections. The energy difference coefficient, by comparing the matching of electrical energy consumption and mechanical energy output between adjacent working events, can identify localized areas with higher electrical input consumption but no corresponding mechanical action. Even if this localized process has not yet caused a significant single-signal anomaly, it can still be assigned corresponding attention weights by the model in subsequent attention calculations. Only when a mismatch trend appears between electrical input and mechanical output will this interval receive higher attention through the energy difference coefficient, thus reducing false alarms caused by occasional fluctuations from a single sensor. The energy difference coefficient calculation in this application is based on existing node and edge features, without increasing the number of model parameters or inference time. Furthermore, due to the use of a small-scale 6-node graph structure, the model computation is relatively small, making it suitable for use in actual working conditions.

[0174] In another optional implementation, for a single running cycle, multiple work events are extracted in chronological order to form an event sequence: For the i-th working event, construct the corresponding event feature vector: For the i-th job event, calculate the query vector, key vector, and value vector respectively:

[0175]

[0176]

[0177]

[0178]

[0179] in, Indicates energy conversion rate, Indicates the amplitude of the force signal. Indicates the amplitude of the displacement signal. Indicates the amplitude of the voltage signal. This indicates the amplitude of the current signal. Represents the query vector for the i-th event; The key vector representing the i-th event; This represents the value vector of the i-th event. , , These represent the basic query mapping matrix, the basic key mapping matrix, and the basic value mapping matrix, respectively. , , These represent the energy conversion rate matrix; the mapping matrices mentioned above are all parameters obtained from model training.

[0180] Regarding the current event and events Calculate attention score:

[0181]

[0182] in, Indicates an event Regarding the event Attention score; Indicates the current event The query vector; Representing candidate events The key vector; : Dimensions of the query vector and key vector.

[0183] Softmax normalization is applied to the attention score of the current event:

[0184]

[0185] in, Indicates an event Regarding the event Attention weights; This represents an exponential function.

[0186] The subsequent calculations yield the same graph embedding vector as in the embodiments of this application, and will not be repeated here. This implementation utilizes the global energy conversion rate to change the node characteristics. , , The calculation method involves selecting the most relevant other event information from the entire event sequence for each work event.

[0187] In S43, the classification output layer calculates probability values ​​based on graph embedding vectors.

[0188]

[0189] in, This represents the weight matrix of the classification output layer, which is learned during the training process. This represents the bias vector of the classification output layer, which is learned during the training process. This represents the graph embedding vector input to the classification output layer. P represents the fault probability value; This represents the Sigmoid activation function (S-shaped growth curve function, its mathematical expression is: It can map any real number to the open interval (0,1). If the fault probability value is greater than or equal to 0.75, the fault detection result is confirmed as faulty; otherwise, it is considered faultless.

[0190] In S40, the deep learning model can be trained in the following way:

[0191] Multiple historical multi-source time-series signals and corresponding historical actual fault results are acquired. These signals include historical voltage, current, displacement, and force signals generated during the equipment's historical operation. Based on these signals, a historical graph structure is constructed. Nodes of the graph represent working events within the operating cycle of the elevator drum brake. Edges represent the state change process between two working events. The historical node features include the historical occurrence time of each working event and the corresponding historical signal amplitude, which includes the amplitude of the historical voltage signal and the amplitude of the historical current signal. The historical graph structure includes at least one of the amplitudes of historical displacement signals and historical force signals. The historical edge features of the historical graph structure include a first historical difference between the historical occurrence times of two working events and a second historical difference between the historical signal amplitudes at the historical occurrence times of the two working events. Based on the historical graph structure, a historical energy difference coefficient is obtained through the historical changes in electrical energy and mechanical energy between the two nodes. The historical graph structure, the historical energy difference coefficient, and the historical real fault results are integrated into training data. A deep learning model is trained and initialized based on the training data so that after the graph structure and the energy difference coefficient are input into the deep learning model, the deep learning model outputs the fault probability value.

[0192] In this embodiment, the collected training dataset is divided into a training set and a test set in an 8:2 ratio. 80% of the data is used for model training, and the remaining 20% ​​is used for testing.

[0193] Among them, the first in the training set Each training sample can be represented as:

[0194]

[0195] in, This represents the nth training sample; This represents the historical graph structure constructed in the nth historical single running cycle. This represents the historical energy difference coefficient corresponding to each edge in the nth historical graph structure. This represents the nth historical real fault label. The historical real fault label is a binary label corresponding to the historical real fault result. . The corresponding historical fault results were confirmed as no fault. The corresponding historical fault results confirm that there is a fault.

[0196] In the training samples of this embodiment, the historical multi-source time-series signals are the signal values ​​obtained by the data acquisition module within a single operating cycle during the device's historical operation. The types and numbers of working events used to construct the nodes in the historical graph structure are the same as D1-D6 in this embodiment. The edges are constructed in the same way as above, and will not be repeated here. The historical node features and historical edge features of the historical graph structure are all derived from historical multi-source time-series signals. The calculation and construction methods for historical node features and historical edge features are the same as those for the graph structure in this embodiment, and will not be repeated here. The historical energy difference coefficient is derived from the historical multi-source time-series signals and the historical graph structure. Its calculation method is the same as that for the energy difference coefficient in this embodiment, and will not be repeated here either.

[0197] The loss function used for training is as follows:

[0198]

[0199] in, Let M represent the loss function, and M represent the number of samples participating in this training calculation. : The historical true fault label of the nth sample ; This represents the probability of the model's fault prediction for the nth sample.

[0200] In this embodiment, the historical graph structure and the historical energy difference coefficient are used together as input features, and historical real-world fault results are used as supervision signals to construct a training dataset. After initializing the deep learning model, a loss function is used for training. After training is complete, during the testing phase, only the graph structure and the energy difference coefficient need to be input, and the model can directly output the fault probability value.

[0201] This application provides an electronic device that may include a processing module and a memory. The memory stores a computer program, which, when executed by the processor, enables the electronic device to perform the corresponding steps in the fault detection method based on multi-source timing signals described above.

[0202] In this embodiment, the processor can be an integrated circuit chip with signal processing capabilities. For example, the processor can be a central processing unit (CPU), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application.

[0203] The memory can be, but is not limited to, random access memory, read-only memory, programmable read-only memory, erasable programmable read-only memory, electrically erasable programmable read-only memory, etc.

[0204] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the electronic device described above can be referred to the corresponding steps in the aforementioned method, and will not be elaborated further here.

[0205] This application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program that, when run on a computer, causes the computer to perform the fault detection method based on multi-source timing signals as described in the above embodiments.

[0206] Computer-readable storage media may be magnetic disks, optical disks, read-only memory, random access memory, flash memory, USB flash drives, hard disks, or solid-state drives, etc., and may also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implement the methods shown in the above embodiments.

[0207] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the aforementioned fault detection method based on multi-source timing signals. The computer program product may exist in a computer-readable storage medium in forms including, but not limited to, source files, executable files, and installation package files.

[0208] Based on the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by hardware or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application can be embodied in the form of a software product. This software product can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as a personal computer, electronic device, or network device, etc.) to execute the methods described in the various implementation scenarios of this application.

[0209] In the embodiments provided in this application, it should be understood that the disclosed methods can also be implemented in other ways. The method embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, program segment, or part of code, which includes one or more executable instructions for implementing a specified logical function. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions. Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0210] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for fault detection based on multi-source time series signals, characterized in that, The method includes: Acquire multi-source timing signals of the device under test within a single operating cycle. The multi-source timing signals include voltage signals, current signals, displacement signals, and force signals during the operation of the device under test. Based on the multi-source time-series signals, a graph structure is constructed. The nodes of the graph structure are the working events within the operating cycle of the device under test. The edges of the graph structure represent the state change process between two working events. The node features include the occurrence time of the working event and the signal amplitude corresponding to the occurrence time. The signal amplitude includes at least one of the amplitude of a voltage signal, the amplitude of a current signal, the amplitude of a displacement signal, and the amplitude of a force signal. The edge features include a first difference between the occurrence times of two working events and a second difference between the signal amplitudes at the occurrence times of two working events. Based on the graph structure, an energy difference coefficient is obtained by measuring the changes in electrical energy and mechanical energy between the two nodes. The energy difference coefficient is used to characterize the degree of energy difference between the electrical energy consumed and the mechanical energy generated by the device under test between the two working events. Both the changes in electrical energy and the changes in mechanical energy are obtained based on the node characteristics of the graph structure. The graph structure and the energy difference coefficient are input into a deep learning model. Based on the graph structure and the energy difference coefficient, the deep learning model obtains a fault probability value, which is used to determine the fault status of the device under test.

2. The method according to claim 1, characterized in that, Based on the graph structure, the energy difference coefficient is obtained by measuring the changes in electrical energy and mechanical energy between the two nodes, including: The change in electrical energy is obtained based on the amplitude of the voltage signal and the amplitude of the current signal in the node characteristics. The change in mechanical energy is obtained based on the amplitude of the force signal and the amplitude of the displacement signal in the node features. Based on the changes in electrical energy and mechanical energy, an energy difference coefficient is obtained according to the energy conversion rate of the device under test. The energy conversion rate is obtained based on the multi-source time-series signal.

3. The method according to claim 1, characterized in that, The deep learning model includes a graph Transformer network, which comprises a graph attention layer, a readout layer, and a classification output layer. Based on the graph structure and the energy difference coefficient, the deep learning model obtains a fault probability value, including: The graph structure and the energy difference coefficient are input into the graph attention layer, and the deep node features are obtained through the graph attention layer. The deep node features are input to the readout layer, and the readout layer concatenates the deep node features to obtain a graph embedding vector; The graph embedding vector is input into the classification output layer, and the fault probability value is obtained through the classification output layer. The fault probability value is used to determine the fault detection result.

4. The method according to claim 3, characterized in that, The process of obtaining deep node features through the graph attention layer includes: Based on the edge features, each feature in the edge features is enhanced using the energy difference coefficient to obtain enhanced edge features. The edge features include at least one feature, which is obtained by the first difference and the second difference. The node features and the enhanced edge features are input into the graph attention layer, and the graph attention layer is used to obtain the first attention score of the node to itself and the second attention score of the node to its neighboring nodes. Normalize the first attention score and the second attention score to obtain the first attention weight of the node to itself and the second attention weight of the node to its neighboring nodes; Based on the first attention weight and the second attention weight, the deep node features of the node are obtained through weighted aggregation.

5. The method according to claim 2, characterized in that, The energy conversion efficiency is obtained based on the multi-source time-series signal, including: Based on the voltage signal and the current signal in the multi-source time-series signal, the electrical energy consumed by the device under test in a single operating cycle is obtained by integrating the voltage signal and the current signal over a single operating cycle. Based on the force signal and displacement signal in the multi-source time-series signal, the mechanical energy generated by the device under test in a single operating cycle is obtained by integrating the force signal and displacement signal over time in a single operating cycle. The energy conversion rate is obtained based on the electrical energy and the mechanical energy.

6. The method according to claim 1, characterized in that, Before acquiring the multi-source timing signals of the device under test within a single operating cycle, the method further includes: Acquire multiple historical multi-source time-series signals and the corresponding historical real fault results of the historical multi-source time-series signals. The historical multi-source time-series signals include historical voltage signals, historical current signals, historical displacement signals, and historical force signals of the equipment during historical operation. Based on the historical multi-source time-series signals, a historical graph structure is constructed. The nodes of the historical graph structure are the working events within the operating cycle of the device under test. The edges of the historical graph structure represent the state change process between two working events. The historical node features of the historical graph structure include the historical occurrence time of the working event and the historical signal amplitude corresponding to the historical occurrence time. The historical signal amplitude includes at least one of the amplitude of historical voltage signal, historical current signal, historical displacement signal, and historical force signal. The historical edge features of the historical graph structure include the historical first difference corresponding to the historical occurrence time of two working events and the historical second difference between the historical signal amplitudes of the two working events at the historical occurrence time. Based on the historical graph structure, the historical energy difference coefficient is obtained by using the historical changes in electrical energy and mechanical energy between the two nodes. The historical graph structure, the historical energy difference coefficient, and the historical real fault results are integrated into training data. The deep learning model is trained and initialized based on the training data, so that after the graph structure and the energy difference coefficient are input into the deep learning model, the deep learning model outputs the fault probability value.

7. An electronic device, characterized in that, The electronic device includes a processor and a memory coupled together, the memory storing a computer program that, when executed by the processor, causes the electronic device to perform the method as described in any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when run on a computer, causes the computer to perform the method as described in any one of claims 1 to 6.

9. A program product, characterized in that, It includes a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 6.

Images