A method for detecting gearbox temperature based on a temperature sensor
By constructing a set of temperature nodes and a thermal causality graph for the gearbox, and combining it with a non-equilibrium hierarchical Kantorovich transport model, the problem of anomaly identification and region localization in gearbox temperature detection in existing technologies is solved, and accurate detection of heat propagation structures and identification of anomaly mechanisms are achieved.
Patent Information
- Application Number
- CN202611102522.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-23
- Publication Date
- 2026-11-17
AI Technical Summary
Existing gearbox temperature detection technologies struggle to uniformly model the sequential response relationships, heat transfer correlations, and heat regeneration relationships among multiple temperature measurement locations. This makes it difficult to accurately identify abnormal mechanisms in the heat transfer structure, leading to false alarms, missed alarms, and insufficient anomaly interpretation.
A non-equilibrium hierarchical Kantorovich transport method with thermal causality constraints is adopted. By constructing a set of temperature nodes and generating a thermal causality graph, the baseline information and mismatch information of heat propagation are extracted. The non-equilibrium hierarchical Kantorovich transport model is used to determine anomalies and identify heat source enhancement type, diffusion hindrance type, regenerative turbulence type and sensor drift type anomalies.
It enables accurate identification and localization of abnormal gearbox temperatures, improves detection sensitivity and structural representation capabilities, reduces the risk of misjudgment, and enhances the reliability and interpretability of detection.
Smart Images

Figure CN123111503A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of temperature detection technology, and in particular to a method for detecting gearbox temperature based on a temperature sensor. Background Technology
[0002] Most existing transmission temperature detection technologies employ single-point or a small number of temperature sensors to collect real-time data from the housing, fluid, or adjacent components. These data are then combined with threshold alarms, temperature compensation, empirical rules, or simplified thermal models to determine the current temperature state. Some solutions correct the detected values based on historical temperatures, ambient temperatures, or thermal balance relationships to improve the accuracy of current temperature estimation. Other solutions establish local temperature rise models or oil temperature prediction models to provide early warnings of abnormal temperature rises. These technologies have certain application value in overheat protection, basic condition monitoring, and general fault early warning, and are widely used in vehicle transmission systems, dual-clutch systems, and gearbox oil temperature monitoring scenarios.
[0003] However, existing technologies mostly focus on detecting single temperature values, local temperature differences, or simple residuals, lacking unified modeling of the sequential response relationships, heat transfer correlations, and heat recovery relationships among multiple temperature measurement locations. This makes it difficult to characterize the structural changes in the heat propagation chain within the gearbox. When local heat source enhancement, heat propagation hindrance, abnormal cross-chain heat transfer, or sensor drift occurs, existing solutions typically only provide conclusions of temperature anomalies or temperature rise deviations, failing to further distinguish the anomaly mechanism or locate the dominant anomaly region. This easily leads to false alarms, missed alarms, and insufficient anomaly interpretation capabilities. Therefore, it is still necessary to propose a gearbox temperature detection method capable of performing layered judgments based on the heat propagation structure.
[0004] Therefore, how to provide a gearbox temperature detection method based on a temperature sensor is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] One objective of this invention is to propose a gearbox temperature detection method based on a temperature sensor. This invention employs a non-equilibrium hierarchical Kantorovich transport method with thermal causality constraints to achieve gearbox temperature anomaly detection and region localization, possessing the advantages of accurate identification, clear classification, and low false positive rate.
[0006] A method for detecting gearbox temperature based on a temperature sensor according to an embodiment of the present invention includes the following steps:
[0007] Collect temperature sequences from various fixed temperature measurement locations in the gearbox and construct a set of temperature nodes;
[0008] A heat cause-effect graph is generated based on the sequential response relationship and heat transfer correlation of each node in the temperature node set.
[0009] The detection windows are divided according to the continuous sampling period. The heat propagation baseline information of each detection window in the healthy operation phase is extracted based on the heat causality diagram to generate a baseline feature group.
[0010] Based on the temperature sequence and thermal cause-effect graph of the current detection window, thermal mismatch information characterizing the current thermal propagation deviation state is extracted, and thermal mismatch feature groups are generated.
[0011] Input the baseline feature set and the thermal mismatch feature set into the non-equilibrium hierarchical Kantorovich transport model with thermal causality constraints to generate transport determination results;
[0012] Based on the transport determination results, determine whether the anomaly is heat source enhancement type, diffusion hindrance type, regenerative turbulence type, or sensor drift type, and output the anomaly level and dominant anomaly area.
[0013] Optionally, the generation of the temperature node set specifically includes:
[0014] Fixed temperature measurement positions are set in the input housing, output housing, front bearing adjacent area, rear bearing adjacent area, oil return adjacent area, oil pan area and housing side wall area of the gearbox, each corresponding to a unique node identifier, forming a candidate temperature node list;
[0015] The sampled temperature at each fixed temperature measurement location in the candidate temperature node list is continuously sampled according to a uniform sampling period, and the sampling results of each fixed temperature measurement location are sorted in chronological order to form the original temperature sequence group.
[0016] Time alignment, outlier removal, and continuity verification are performed on the original temperature sequence group to retain valid temperature segments that meet the requirements of a unified time base and continuous sampling, thus forming a regular temperature sequence group.
[0017] Each effective temperature segment in the regular temperature sequence group is bound to its corresponding node identifier. Each node identifier corresponds to a unique regular temperature sequence, forming a temperature node.
[0018] Temperature nodes are selected and encapsulated based on their fixed positions on the gearbox and the continuous effectiveness of the regular temperature sequence, thus generating a set of temperature nodes.
[0019] Optionally, the generation of the heat causality graph specifically includes:
[0020] Taking the regular temperature sequence corresponding to each temperature node in the temperature node set as the object, the arrival order of temperature changes of each temperature node is compared one by one under a unified time reference. The upstream and downstream nodes and their corresponding sequential relationships of each node pair in each comparison period are recorded to form a node sequential response record set.
[0021] Perform stability screening on the node response record set, remove node pairs whose arrival order is repeatedly reversed, and retain node pairs whose arrival order remains consistent over multiple consecutive comparison periods to form a candidate response relationship set;
[0022] A heat transfer correlation record set is formed by checking the continuous following state, stable decay state and heat recovery corresponding state between the nodes in the candidate response relationship set.
[0023] For candidate heat transfer node pairs in the heat transfer association record set, extract the propagation time delay base value, attenuation relationship and heat recovery association to form edge attribute groups, and bind them with the corresponding candidate heat transfer nodes to form heat propagation edges;
[0024] According to the connection direction of the upstream and downstream nodes of each heat propagation edge, heat propagation edges with continuous transmission relationship are sequentially spliced together to form a set of heat propagation chains;
[0025] Using the set of temperature nodes as graph nodes, the set of heat propagation edges as graph edges, and the set of heat propagation chains as link constraints, the propagation time delay base value, attenuation relationship and heat recovery association in the edge attribute group are written into the corresponding graph edges to generate a thermal cause-effect graph.
[0026] Optionally, the edge attribute group is obtained by summarizing the delay duration, attenuation ratio and heat recovery correspondence ratio of candidate heat transfer node pairs within the effective comparison period. For each candidate heat transfer node pair, the comparison period that simultaneously satisfies the conditions of stable order, continuous following, stable attenuation and heat recovery correspondence is retained as the effective comparison period. The propagation time delay base value, attenuation relationship and heat recovery association are calculated respectively. The same candidate heat transfer node pair and its corresponding propagation time delay base value, attenuation relationship and heat recovery association are encapsulated as heat propagation edge.
[0027] Optionally, the generation of the baseline feature set specifically includes:
[0028] In the regular temperature sequence corresponding to the set of temperature nodes, a continuous sampling period with continuous and effective temperature nodes, stable node response relationship and stable heat transfer correlation is selected as the healthy operation stage. The healthy operation stage is divided according to a unified window duration and a unified sliding step size to form a health detection window sequence.
[0029] Around each health detection window, based on the set of heat propagation chains in the heat propagation graph, the connection direction of heat propagation edges, and the base value of propagation time delay, attenuation relationship and heat recovery association of each heat propagation edge, the link order and corresponding node range of each heat propagation chain within the health detection window are determined, forming a link benchmark mapping table.
[0030] According to the link reference mapping table, extract the heat propagation reference quantity of each heat propagation chain in each health detection window to form a window-level heat propagation reference segment.
[0031] Window-level heat propagation reference segments belonging to the same heat propagation chain are merged according to the heat propagation chain identifier. Consistency screening and central summarization are performed on the merged window-level heat propagation reference segments to form chain-level reference entries.
[0032] The chain-level benchmark entries corresponding to each heat propagation chain are encapsulated according to the order of the heat propagation chain set to generate benchmark feature groups.
[0033] Optionally, the generation of the thermal mismatch feature set specifically includes:
[0034] The current continuous sampling period is divided into the current detection window sequence according to the same window duration and sliding step size as the health detection window sequence.
[0035] Based on the set of heat propagation chains, the connection direction of heat propagation edges, and the base value of propagation time delay, attenuation relationship and heat recovery association of each heat propagation edge in the heat cause-effect graph, the link order and corresponding node range of each heat propagation chain in each current detection window sequence are determined to form the current link mapping table.
[0036] According to the current link mapping table, extract the current heat propagation amount of each heat propagation chain in each current detection window to form a window-level current heat propagation segment.
[0037] The window-level current heat propagation segment is aligned with the corresponding chain-level benchmark entry in the benchmark feature group according to the heat propagation chain identifier, heat propagation edge order and node range. The time delay deviation result, attenuation deviation result, heat recovery deviation result and node temperature difference deviation result are extracted to form a window-level heat mismatch segment.
[0038] The window-level thermal mismatch segments corresponding to each heat propagation chain within the same current detection window are merged and sequentially encapsulated to form the thermal mismatch information corresponding to the current detection window.
[0039] The thermal mismatch information corresponding to each current detection window in the current detection window sequence is encapsulated according to the order of the detection windows to generate thermal mismatch feature groups.
[0040] Optionally, the generation of the transport determination result specifically includes:
[0041] A non-equilibrium hierarchical Kantorovich transport model with thermal causality constraints is constructed around the baseline feature set, the thermal mismatch feature set, and the thermal causality graph.
[0042] According to the heat propagation chain set, the baseline feature group and the thermal mismatch feature group are paired in the same window and split in the same chain, and the pairing results are written into the intra-chain transport layer to form an intra-chain transport unit set.
[0043] Inter-chain relationships are constructed around the set of intra-chain transport units and written into the inter-chain transport layer to form a hierarchical transport structure;
[0044] Constraints are configured on the hierarchical transport structure in thermal causality constraint units and non-equilibrium constraint units to form a constraint transport cost table;
[0045] In the intra-chain transport layer, position matching and cost accumulation are performed according to the constraint transport cost table to obtain the intra-chain transport result;
[0046] In the inter-chain transport layer, cross-chain transfers and reheat checks are performed based on the constraint transport cost table to obtain the inter-chain transport results.
[0047] According to the current detection window order, the intra-chain transport results and inter-chain transport results corresponding to each heat propagation chain are merged to generate transport judgment results.
[0048] Optionally, the non-equilibrium hierarchical Kantorovich transport model with thermal causal constraints uses a reference feature set as the reference side input, a thermal mismatch feature set as the mismatch side input, and a thermal causal graph as the link constraint input, including an intra-link transport layer, an inter-link transport layer, thermal causal constraint units, and non-equilibrium constraint units.
[0049] The intra-chain transport layer is used to perform position matching and cost accumulation between the reference side and the mismatch side within the same heat propagation chain. The inter-chain transport layer is used to perform cross-chain matching and cost accumulation for mismatch transfer between different heat propagation chains. The thermal causality constraint unit is used to constrain the transport direction, propagation time delay, attenuation relationship and regeneration correlation. The non-equilibrium constraint unit is used to retain and accumulate the mismatch and missing parts that cannot be matched.
[0050] Optionally, the generation of the anomaly level and the dominant anomaly region specifically includes:
[0051] Extract the transport judgment results corresponding to each current detection window in the current detection window order, and extract the intra-chain transport cost, inter-chain transport cost, new mismatch, missing mismatch, node dominance distribution and heat recovery imbalance cost for each transport judgment result to form an anomaly judgment sequence.
[0052] Based on the anomaly determination sequence, anomaly type determination is performed on each current detection window to form a window-level anomaly type result;
[0053] Perform continuity filtering and stability filtering on window-level anomaly type results to generate anomaly type results;
[0054] Based on the anomaly type results, an anomaly level determination is performed on each current detection window to generate an anomaly level result;
[0055] Extract the dominant anomaly region based on the anomaly type and anomaly level results.
[0056] The beneficial effects of this invention are:
[0057] This invention constructs a set of temperature nodes at multiple fixed temperature measurement locations in the gearbox and generates a thermal causality graph based on the sequential response relationship and heat transfer correlation between the temperature nodes. This elevates the temperature detection object from the single-point temperature value or local temperature rise in existing technologies to a thermal propagation structure capable of characterizing propagation time delay, attenuation relationship, and regeneration correlation. Based on this thermal causality graph, it further extracts thermal propagation baseline information during the healthy operation phase and thermal mismatch information of the current detection window. Furthermore, it uses a non-equilibrium hierarchical Kantorovich transport model with thermal causality constraints to uniformly determine intra-chain transport, inter-chain transport, newly added mismatch, missing mismatch, and regeneration imbalance state. This enables more accurate identification of local thermal propagation deviations, cross-chain thermal transfer mismatches, and regeneration region imbalances, improving the detection sensitivity and structural representation capability for complex thermal anomalies.
[0058] After obtaining the transport determination result, this invention further combines intra-chain transport cost, inter-chain transport cost, newly added mismatch amount, missing mismatch amount, node dominance distribution, and regenerative imbalance cost to distinguish between heat source enhancement anomalies, diffusion hindrance anomalies, regenerative disorder anomalies, and sensor drift anomalies, and outputs the anomaly level and dominant anomaly region. This not only enables the detection of gearbox temperature anomalies but also the identification of anomaly mechanisms and the location of anomaly regions. Compared with existing technologies that rely solely on threshold comparison, temperature compensation, or single residual judgment, this invention effectively reduces the risk of misjudgment caused by short-term fluctuations, local disturbances, or single-point drift, enhances the separability of different anomaly types and the interpretability of results, thereby improving the accuracy, reliability, and engineering application value of gearbox temperature detection. Attached Figure Description
[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0060] Figure 1 This is a flowchart of a gearbox temperature detection method based on a temperature sensor proposed in this invention;
[0061] Figure 2 This is a flowchart illustrating the construction of a thermal cause-effect graph for a gearbox temperature detection method based on a temperature sensor proposed in this invention.
[0062] Figure 3 This is a flowchart of the layered transport anomaly determination method of a gearbox temperature detection method based on a temperature sensor proposed in this invention. Detailed Implementation
[0063] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0064] refer to Figures 1-3 A method for detecting gearbox temperature based on a temperature sensor includes the following steps:
[0065] Collect temperature sequences from various fixed temperature measurement locations in the gearbox and construct a set of temperature nodes;
[0066] A heat cause-effect graph is generated based on the sequential response relationship and heat transfer correlation of each node in the temperature node set.
[0067] The detection windows are divided according to the continuous sampling period. The heat propagation baseline information of each detection window in the healthy operation phase is extracted based on the heat causality diagram to generate a baseline feature group.
[0068] Based on the temperature sequence and thermal cause-effect graph of the current detection window, thermal mismatch information characterizing the current thermal propagation deviation state is extracted, and thermal mismatch feature groups are generated.
[0069] Input the baseline feature set and the thermal mismatch feature set into the non-equilibrium hierarchical Kantorovich transport model with thermal causality constraints to generate transport determination results;
[0070] Based on the transport determination results, determine whether the anomaly is heat source enhancement type, diffusion hindrance type, regenerative turbulence type, or sensor drift type, and output the anomaly level and dominant anomaly area.
[0071] In this embodiment, the generation of the temperature node set specifically includes:
[0072] Fixed temperature measurement positions are set in the input housing, output housing, front bearing adjacent area, rear bearing adjacent area, oil return adjacent area, oil pan area and housing side wall area of the gearbox, each corresponding to a unique node identifier, forming a candidate temperature node list;
[0073] The sampled temperature at each fixed temperature measurement location in the candidate temperature node list is continuously sampled according to a uniform sampling period, and the sampling results of each fixed temperature measurement location are sorted in chronological order to form the original temperature sequence group.
[0074] Time alignment, outlier removal, and continuity verification are performed on the original temperature sequence group to retain valid temperature segments that meet the requirements of a unified time base and continuous sampling, thus forming a regular temperature sequence group.
[0075] Each effective temperature segment in the regular temperature sequence group is bound to its corresponding node identifier. Each node identifier corresponds to a unique regular temperature sequence, forming a temperature node.
[0076] Temperature nodes are selected and encapsulated based on their fixed positions on the gearbox and the continuous effectiveness of the regular temperature sequence, thus generating a set of temperature nodes.
[0077] In this embodiment, the generation of the heat cause-effect graph specifically includes:
[0078] Taking the regular temperature sequence corresponding to each temperature node in the temperature node set as the object, the arrival order of temperature changes of each temperature node is compared one by one under a unified time reference. The upstream and downstream nodes and their corresponding sequential relationships of each node pair in each comparison period are recorded to form a node sequential response record set.
[0079] Perform stability screening on the node response record set, remove node pairs whose arrival order is repeatedly reversed, and retain node pairs whose arrival order remains consistent over multiple consecutive comparison periods to form a candidate response relationship set;
[0080] A heat transfer correlation record set is formed by checking the continuous following state, stable decay state and heat recovery corresponding state between the nodes in the candidate response relationship set.
[0081] The verification of the continuous following state, stable decay state, and regeneration corresponding state includes: for each node pair in the candidate sequential response relationship set, extracting the starting time of upstream node temperature change, the starting time of downstream node temperature change, upstream temperature difference amplitude, downstream temperature difference amplitude, same-direction temperature difference amplitude in the oil return adjacent area, same-direction temperature difference amplitude in the oil pan area, and same-direction temperature difference amplitude in the box side wall area within each comparison period; first, subtracting the starting time of upstream node temperature change from the starting time of downstream node temperature change to obtain the delay time of the comparison period; then, dividing the downstream temperature difference amplitude by the upstream temperature difference amplitude to obtain the decay ratio of the comparison period; and finally, dividing the sum of the same-direction temperature difference amplitudes in the oil return adjacent area, the oil pan area, and the box side wall area by three times the downstream temperature difference amplitude to obtain the regeneration corresponding ratio of the comparison period.
[0082] The median value of the delay duration is determined by sorting all comparison periods by their delay durations in descending order. The median value of the attenuation ratio is determined by sorting all comparison periods by their attenuation ratios in descending order. The median value of the heat recovery correspondence ratio is determined by sorting all comparison periods by their heat recovery correspondence ratios in descending order. The synchronization degree is determined by the proportion of comparison periods for which the absolute value of the difference between the delay duration and the median delay duration is not greater than the time delay deviation threshold to the total number of comparison periods for that node. The attenuation consistency degree is determined by the proportion of comparison periods for which the absolute value of the difference between the attenuation ratio and the median attenuation ratio is not greater than the attenuation deviation threshold to the total number of comparison periods for that node. The heat recovery correspondence degree is determined by the proportion of comparison periods for which the absolute value of the difference between the heat recovery correspondence ratio and the median heat recovery correspondence ratio is not greater than the heat recovery deviation threshold to the total number of comparison periods for that node.
[0083] When the stable sequence of the same node pair is established, and the synchronization degree is not lower than the synchronization degree screening threshold, the attenuation consistency degree is not lower than the attenuation consistency screening threshold, and the heat recovery correspondence degree is not lower than the heat recovery correspondence screening threshold, the node pair is recorded as a candidate heat transfer node pair and written into the heat transfer association record set.
[0084] For candidate heat transfer node pairs in the heat transfer association record set, extract the propagation time delay base value, attenuation relationship and heat recovery association to form edge attribute groups, and bind them with the corresponding candidate heat transfer nodes to form heat propagation edges;
[0085] According to the connection direction of the upstream and downstream nodes of each heat propagation edge, heat propagation edges with continuous transmission relationship are sequentially spliced together to form a set of heat propagation chains;
[0086] Using the set of temperature nodes as graph nodes, the set of heat propagation edges as graph edges, and the set of heat propagation chains as link constraints, the propagation time delay base value, attenuation relationship and heat recovery association in the edge attribute group are written into the corresponding graph edges to generate a thermal cause-effect graph.
[0087] In this embodiment, the edge attribute group is obtained by summarizing the delay time, attenuation ratio and heat recovery correspondence ratio of candidate heat transfer node pairs within the effective comparison period. For each candidate heat transfer node pair, the comparison period that simultaneously satisfies the conditions of stable order, continuous following, stable attenuation and heat recovery correspondence is retained as the effective comparison period. The propagation time delay base value, attenuation relationship and heat recovery association are calculated respectively. The same candidate heat transfer node pair and its corresponding propagation time delay base value, attenuation relationship and heat recovery association are encapsulated as heat propagation edge.
[0088] Among them, the propagation time delay base value is obtained by sorting the delay time of all effective comparison periods of the candidate heat transfer node in order of magnitude; the attenuation relationship is obtained by sorting the attenuation ratio of all effective comparison periods of the candidate heat transfer node in order of magnitude; and the heat recovery association is obtained by sorting the heat recovery ratio of all effective comparison periods of the candidate heat transfer node in order of magnitude.
[0089] In this embodiment, the generation of the baseline feature set specifically includes:
[0090] In the regular temperature sequence corresponding to the set of temperature nodes, a continuous sampling period with continuous and effective temperature nodes, stable node response relationship and stable heat transfer correlation is selected as the healthy operation stage. The healthy operation stage is divided according to a unified window duration and a unified sliding step size to form a health detection window sequence.
[0091] Around each health detection window, based on the set of heat propagation chains in the heat propagation graph, the connection direction of heat propagation edges, and the base value of propagation time delay, attenuation relationship and heat recovery association of each heat propagation edge, the link order and corresponding node range of each heat propagation chain within the health detection window are determined, forming a link benchmark mapping table.
[0092] According to the link reference mapping table, extract the heat propagation reference quantity of each heat propagation chain in each health detection window to form a window-level heat propagation reference segment.
[0093] The extraction of the intra-chain heat propagation benchmark quantity involves checking the actual delay time, actual attenuation result, and actual heat recovery result between upstream and downstream nodes within the health detection window along the connection direction of the heat propagation edge of the same heat propagation chain. These results are then aligned with the propagation time delay base value, attenuation relationship, and heat recovery correlation of the corresponding heat propagation edge. The actual delay time corresponding to each heat propagation edge within the same heat propagation chain is arranged sequentially according to the connection direction to form a time delay benchmark segment. The actual attenuation result corresponding to each heat propagation edge within the same heat propagation chain is arranged sequentially according to the connection direction to form an attenuation benchmark segment. The actual heat recovery result corresponding to each heat propagation edge within the same heat propagation chain is arranged sequentially according to the connection direction to form a heat recovery benchmark segment. The node temperature difference status corresponding to each temperature node within the health detection window is summarized according to the node range of the heat propagation chain to form a node benchmark segment. Finally, the time delay benchmark segment, attenuation benchmark segment, heat recovery benchmark segment, and node benchmark segment are merged into a window-level heat propagation benchmark segment corresponding to the health detection window.
[0094] Window-level heat propagation reference segments belonging to the same heat propagation chain are merged according to the heat propagation chain identifier. Consistency screening and central summarization are performed on the merged window-level heat propagation reference segments to form chain-level reference entries.
[0095] The consistency screening and central aggregation process involves checking whether the heat propagation edge order and node range are consistent for all window-level heat propagation reference segments corresponding to the same heat propagation chain. Then, it checks whether the time delay reference segment, attenuation reference segment, regeneration reference segment and node reference segment in each window-level heat propagation reference segment are within the preset fluctuation range. Window-level heat propagation reference segments with consistent heat propagation edge order, consistent node range and fluctuation of each segment not exceeding the corresponding preset fluctuation range are retained. The retained time delay reference segment, attenuation reference segment, regeneration reference segment and node reference segment are aligned in position and aggregated according to the median value of the corresponding position to form a chain-level reference entry that corresponds one-to-one with the heat propagation chain.
[0096] The chain-level benchmark entries corresponding to each heat propagation chain are encapsulated according to the order of the heat propagation chain set to generate benchmark feature groups.
[0097] In this embodiment, the generation of the thermal mismatch feature set specifically includes:
[0098] The current continuous sampling period is divided into the current detection window sequence according to the same window duration and sliding step size as the health detection window sequence.
[0099] Based on the set of heat propagation chains, the connection direction of heat propagation edges, and the base value of propagation time delay, attenuation relationship and heat recovery association of each heat propagation edge in the heat cause-effect graph, the link order and corresponding node range of each heat propagation chain in each current detection window sequence are determined to form the current link mapping table.
[0100] According to the current link mapping table, extract the current heat propagation amount of each heat propagation chain in each current detection window to form a window-level current heat propagation segment.
[0101] The extraction of the current heat propagation within the chain is achieved by checking the actual delay time, actual attenuation result, and actual heat recovery result between upstream and downstream nodes within the current detection window along the connection direction of the heat propagation edge of the same heat propagation chain. The actual delay time corresponding to each heat propagation edge within the same heat propagation chain is arranged sequentially according to the connection direction to form the current time delay segment. The actual attenuation result corresponding to each heat propagation edge within the same heat propagation chain is arranged sequentially according to the connection direction to form the current attenuation segment. The actual heat recovery result corresponding to each heat propagation edge within the same heat propagation chain is arranged sequentially according to the connection direction to form the current heat recovery segment. The node temperature difference status corresponding to each temperature node within the current detection window is summarized according to the node range of the heat propagation chain to form the current node segment. The current time delay segment, current attenuation segment, current heat recovery segment, and current node segment are merged into the window-level current heat propagation segment corresponding to the current detection window.
[0102] The window-level current heat propagation segment is aligned with the corresponding chain-level benchmark entry in the benchmark feature group according to the heat propagation chain identifier, heat propagation edge order and node range. The time delay deviation result, attenuation deviation result, heat recovery deviation result and node temperature difference deviation result are extracted to form a window-level heat mismatch segment.
[0103] The position alignment process first checks whether the current heat propagation segment at the window level is consistent with the heat propagation edge order and node range of the corresponding chain-level reference entry. Then, it compares the current time-delay segment with the time-delay reference segment in the corresponding chain-level reference entry position by position to obtain the time-delay deviation result. The current attenuation segment is compared with the attenuation reference segment in the corresponding chain-level reference entry position by position to obtain the attenuation deviation result. The current heat recovery segment is compared with the heat recovery reference segment in the corresponding chain-level reference entry position by position to obtain the heat recovery deviation result. The current node segment is compared with the node reference segment in the corresponding chain-level reference entry position by position to obtain the node temperature difference deviation result. For the same heat propagation chain, the time-delay deviation result, attenuation deviation result, heat recovery deviation result, and node temperature difference deviation result are combined into a window-level heat mismatch segment according to their original position order.
[0104] The window-level thermal mismatch segments corresponding to each heat propagation chain within the same current detection window are merged and sequentially encapsulated to form the thermal mismatch information corresponding to the current detection window.
[0105] The merging and sequential encapsulation process involves sequentially splicing the window-level thermal mismatch segments corresponding to each thermal propagation chain according to the arrangement order of the thermal propagation chain set, and retaining the thermal propagation chain identifier, thermal propagation edge order and node range corresponding to each window-level thermal mismatch segment, so that the time delay deviation results, attenuation deviation results, heat recovery deviation results and node temperature difference deviation results in the current detection window are aggregated into thermal mismatch information under the same sequential framework.
[0106] The thermal mismatch information corresponding to each current detection window in the current detection window sequence is encapsulated according to the order of the detection windows to generate thermal mismatch feature groups.
[0107] In this embodiment, the generation of the transportation determination result specifically includes:
[0108] A non-equilibrium hierarchical Kantorovich transport model with thermal causality constraints is constructed around the baseline feature set, the thermal mismatch feature set, and the thermal causality graph.
[0109] According to the heat propagation chain set, the baseline feature group and the thermal mismatch feature group are paired in the same window and split in the same chain, and the pairing results are written into the intra-chain transport layer to form an intra-chain transport unit set.
[0110] The same-window pairing and same-chain splitting include extracting the window-level thermal mismatch segments corresponding to each thermal propagation chain from the thermal mismatch information corresponding to each current detection window, extracting the chain-level benchmark entries corresponding to the same thermal propagation chain from the benchmark feature group, pairing them one by one according to the thermal propagation chain identifier, thermal propagation edge order and node range, taking the chain-level benchmark entries in each pairing result as the benchmark side and the corresponding window-level thermal mismatch segments as the mismatch side, forming the intra-chain transport unit of the corresponding thermal propagation chain;
[0111] Inter-chain relationships are constructed around the set of intra-chain transport units and written into the inter-chain transport layer to form a hierarchical transport structure;
[0112] The inter-chain association is obtained by checking whether the node ranges of different heat propagation chains are adjacent, whether they correspond to the same return oil adjacent area, oil pan area or tank side wall area, and whether there is a return heat association intersection of the corresponding heat propagation edges. When at least one of the following conditions is met, such as adjacent node ranges, corresponding to the same return heat area or having a return heat association intersection, an inter-chain association is established between the corresponding heat propagation chains, and the cross-chain transfer path between the corresponding heat propagation chains is written into the inter-chain transport layer. The set of intra-chain transport units in the intra-chain transport layer and the inter-chain association in the inter-chain transport layer are jointly organized into a hierarchical transport structure.
[0113] Constraints are configured on the hierarchical transport structure in thermal causality constraint units and non-equilibrium constraint units to form a constraint transport cost table;
[0114] The thermal causality constraint unit checks the time delay deviation, attenuation deviation, heat recovery deviation, and node temperature difference deviation results at each position on the mismatch side of each intra-chain transport unit in the intra-chain transport layer. It determines whether these results meet the allowable deviation ranges for the heat propagation edge connection direction, propagation time delay base value, attenuation relationship, and heat recovery association. Based on the degree of satisfaction, it configures the direction deviation cost, time delay deviation cost, attenuation deviation cost, and heat recovery deviation cost for the corresponding position. It also checks the cross-chain transfer paths corresponding to the inter-chain association relationships in the inter-chain transport layer. It determines whether these paths meet the link order, node range adjacency relationship, and heat recovery region connectivity relationship in the heat propagation chain set. Based on the degree of correspondence between the cross-chain transfer direction and the heat recovery region, it configures the inter-chain transport cost and writes it into the constraint transport cost table.
[0115] The allowable deviation range is determined according to the historical stable fluctuation of the heat propagation edge in the health detection window sequence. For the propagation time delay base value, attenuation relationship and heat recovery association corresponding to each heat propagation edge, the actual value of the heat propagation edge at the same position in all health detection windows is extracted. The median value at the same position in the corresponding chain-level reference entry is used as the center value. The absolute deviation between the actual value at the same position in each health detection window and the center value is calculated. All absolute deviations are sorted by size and the median value is taken. The corresponding preset tolerance value is superimposed on the median value to obtain the allowable deviation range of propagation time delay, allowable deviation range of attenuation relationship and allowable deviation range of heat recovery association. The allowable deviation range of node temperature difference deviation result is determined according to the node temperature difference status at the same position of the node reference segment in all health detection windows.
[0116] The directional deviation cost, time delay deviation cost, attenuation deviation cost, and regeneration deviation cost are selected from pre-set fixed values based on the consistency between the actual transmission direction and the heat propagation edge connection direction at the corresponding position of the mismatch side, the difference between the actual delay time and the propagation time delay base value, the difference between the actual attenuation result and the attenuation relationship, and the difference between the actual regeneration result and the regeneration correlation. When the corresponding comparison result is within the corresponding allowable deviation range, it is recorded as the basic directional cost, basic time delay cost, basic attenuation cost, and basic regeneration cost, respectively. When the corresponding comparison result exceeds the corresponding allowable deviation range but does not exceed the corresponding allowable deviation range, it is recorded as the basic directional cost, basic time delay cost, basic attenuation cost, and basic regeneration cost. When the deviation is twice the allowable range, it is recorded as the first directional incremental cost, the first time delay incremental cost, the first attenuation incremental cost, and the first regeneration incremental cost, respectively. When the corresponding comparison result exceeds twice the corresponding allowable deviation range, it is recorded as the second directional incremental cost, the second time delay incremental cost, the second attenuation incremental cost, and the second regeneration incremental cost, respectively. Among the directional deviation costs, if the actual transmission direction is opposite to the connection direction of the heat propagation edge but is still within the range of adjacent heat propagation edges in the same heat propagation chain, it is recorded as the first directional incremental cost. If the actual transmission direction is opposite to the connection direction of the heat propagation edge and crosses a non-adjacent heat propagation edge, it is recorded as the second directional incremental cost.
[0117] The deviations of the unbalanced constraint unit that cannot complete the position matching or cross-chain transfer within the corresponding allowable range are recorded as newly added mismatch and missing mismatch, respectively, and written into the constraint transport cost table.
[0118] In the intra-chain transport layer, position matching and cost accumulation are performed according to the constraint transport cost table to obtain the intra-chain transport result;
[0119] The location matching and cost accumulation are achieved by first performing same-location matching on the reference side and the mismatch side according to the heat propagation edge order and node range within the same chain transport unit. For the part that has not completed same-location matching, adjacent matching is performed according to the adjacent heat propagation edge order. During each matching, the corresponding directional deviation cost, time delay deviation cost, attenuation deviation cost, and heat recovery deviation cost in the thermal causality constraint unit are called and accumulated. For the deviation results of the completed matching, the node dominance is accumulated according to the node range to which it belongs. For the part that has not completed matching and is retained on the mismatch side, the non-equilibrium constraint unit is called to accumulate it as the new mismatch amount. For the part that has not completed matching and is retained on the reference side, the non-equilibrium constraint unit is called to accumulate it as the missing mismatch amount. All matching costs in the same chain transport unit are merged with the corresponding new mismatch amount and missing mismatch amount to form the in-chain transport result corresponding to the heat propagation chain.
[0120] The node dominance is obtained by accumulating the cost contribution of the deviation results of the completed match to the corresponding node range. For each deviation result of the completed match, if the deviation result comes from the time delay deviation result, the attenuation deviation result, or the heat recovery deviation result, the direction deviation cost, time delay deviation cost, attenuation deviation cost, and heat recovery deviation cost accumulated in this match are evenly distributed to the upstream node range and downstream node range of the heat propagation edge corresponding to the matching position. If the deviation result comes from the node temperature difference deviation result, all the costs accumulated in this match are recorded in the node range corresponding to the node temperature difference deviation result. All completed match results belonging to the same node range within the same current detection window are accumulated to obtain the node dominance of the node range within the current detection window.
[0121] In the inter-chain transport layer, cross-chain transfers and reheat checks are performed based on the constraint transport cost table to obtain the inter-chain transport results.
[0122] The cross-chain transfer and regeneration verification process takes the newly added mismatches and missing mismatches that have not been matched in the intra-chain transport layer as the cross-chain transfer objects. In the inter-chain transport layer, cross-chain matching is performed according to the inter-chain association, the adjacency of the node range, and the connectivity of the regeneration region. For the part that meets the cross-chain matching conditions, the corresponding inter-chain transport cost in the thermal causal constraint unit is called to accumulate. For the part involving cross-chain transfer between the oil return adjacent area, the oil pan area, and the box side wall area, the degree of deviation of its regeneration association is further checked in the thermal causal constraint unit, and the regeneration imbalance cost is accumulated. For the part that fails to complete the cross-chain matching, the non-balance constraint unit is called to continue to retain the corresponding newly added mismatches or missing mismatches. The inter-chain transport cost, regeneration imbalance cost, and the retained newly added mismatches and missing mismatches corresponding to all cross-chain matching results are merged to form the inter-chain transport result.
[0123] A chain-to-chain relationship is determined to exist between two heat propagation chains when at least two of the following conditions are met: the upstream node range of the latter heat propagation chain and the downstream node range of the former heat propagation chain have the same temperature node or directly adjacent temperature nodes; each of the two heat propagation chains has at least one heat propagation edge corresponding to the same return oil adjacent area, oil pan area, or tank sidewall area; the heat propagation edges of the two heat propagation chains corresponding to the same return area have their heat propagation associations within their respective allowable deviation ranges, and the connection directions of the heat propagation edges of both chains point to the same return area or are both away from the same return area.
[0124] The cost of heat recovery imbalance is determined based on the cumulative heat recovery deviation in the adjacent oil return area, oil pan area, and tank sidewall area involved in the cross-chain transfer process. For each cross-chain transfer path, the difference between the actual heat recovery result in the corresponding heat recovery area and the heat recovery association is first checked to see if it exceeds the allowable deviation range of the heat recovery association. If it does not exceed the allowable deviation range of the heat recovery association, only the basic heat recovery cost is retained and not included in the heat recovery imbalance cost. If it exceeds the allowable deviation range of the heat recovery association and only exceeds the limit in one heat recovery area, the first heat recovery imbalance cost is recorded. If it exceeds the allowable deviation range of the heat recovery association and exceeds the limit in two or more heat recovery areas, or if the cross-chain transfer direction is opposite to the direction of connection of the heat recovery area, the second heat recovery imbalance cost is recorded.
[0125] According to the current detection window order, the intra-chain transport results and inter-chain transport results corresponding to each heat propagation chain are merged to generate transport judgment results.
[0126] In this embodiment, the non-equilibrium hierarchical Kantorovich transport model with thermal causal constraints uses a reference feature set as the reference side input, a thermal mismatch feature set as the mismatch side input, and a thermal causal graph as the link constraint input. It includes an intra-link transport layer, an inter-link transport layer, thermal causal constraint units, and non-equilibrium constraint units.
[0127] The non-equilibrium hierarchical Kantorovich transport model with thermal causality constraints is an improvement on the residual distribution change point and fault detection algorithm based on Kantorovich Distance. Compared with the existing general Kantorovich Distance algorithm, which only calculates the overall distance between the training residual distribution and the test residual distribution, this invention improves the transport object to a thermal propagation structure feature composed of propagation time delay, attenuation relationship, regeneration correlation and node temperature difference state. It improves the free transport space to a thermal propagation chain space constrained by thermal causality graph. It improves the conserved distance comparison to a non-equilibrium transport process that can retain the newly added mismatch and the missing mismatch. Furthermore, it improves the single-layer distance output to a hierarchical transport result of intra-chain transport layer and inter-chain transport layer, so that the transport judgment result can simultaneously reflect the local thermal propagation deviation, cross-chain thermal transfer mismatch and regeneration region imbalance state.
[0128] The intra-chain transport layer is used to perform position matching and cost accumulation between the reference side and the mismatch side within the same heat propagation chain. The inter-chain transport layer is used to perform cross-chain matching and cost accumulation for mismatch transfer between different heat propagation chains. The thermal causality constraint unit is used to constrain the transport direction, propagation time delay, attenuation relationship and regeneration correlation. The non-equilibrium constraint unit is used to retain and accumulate the mismatch and missing parts that cannot be matched.
[0129] The training of the non-equilibrium hierarchical Kantorovich transport model with thermal causality constraints involves constructing a training sample set by extracting the baseline feature set corresponding to the healthy operation phase, the thermal mismatch feature set corresponding to the known abnormal phase, and the abnormality type labeling results. The model parameters are then iteratively updated. The loss function is jointly calculated based on the consistency of the target classification between the intra-chain transport results and the inter-chain transport results corresponding to each detection window in the training sample set, the consistency of the target abnormality type, the balance between the newly added mismatch and the missing mismatch, and the stability of the node dominance distribution. The minimum total loss value is used as the parameter update objective. When the decrease in the total loss value of multiple consecutive training rounds does not exceed the preset convergence threshold, and the abnormality type judgment results in the verification sample set remain unchanged or the fluctuation does not exceed the preset stability threshold, the convergence state is determined, and the current parameters are fixed as transport judgment parameters.
[0130] In this embodiment, the generation of the anomaly level and the dominant anomaly region specifically includes:
[0131] Extract the transport judgment results corresponding to each current detection window in the current detection window order, and extract the intra-chain transport cost, inter-chain transport cost, new mismatch, missing mismatch, node dominance distribution and heat recovery imbalance cost for each transport judgment result to form an anomaly judgment sequence.
[0132] Based on the anomaly determination sequence, anomaly type determination is performed on each current detection window to form a window-level anomaly type result;
[0133] The anomaly type determination includes:
[0134] When the intra-chain transport cost corresponding to the same heat propagation chain is in the high cost range, the newly added mismatch is greater than the missing mismatch, the node dominance distribution is concentrated in the range of upstream nodes of the heat propagation chain, and the inter-chain transport cost and the heat recovery imbalance cost are both in the low cost range, it is judged as a heat source enhancement anomaly.
[0135] When the intra-chain transport cost corresponding to the same heat propagation chain is in the high cost range, the missing mismatch is greater than the new mismatch, the node dominance is concentrated in the upstream node range of the heat propagation chain, and the missing mismatch appears continuously in the downstream node range along the heat propagation edge connection direction, it is judged as a diffusion hindrance anomaly.
[0136] When the inter-chain transport cost is in the high cost range, the heat recovery imbalance cost is in the high cost range, and the node dominance distribution simultaneously covers the node range of at least two heat propagation chains with inter-chain relationships, it is judged as a heat recovery disorder anomaly.
[0137] When the node dominance distribution is concentrated in a single node range, the inter-chain transport cost is in a low cost range, the heat recovery imbalance cost is in a low cost range, and neither the newly added mismatch nor the missing mismatch extends along the adjacent heat propagation edge, it is judged as a sensor drift anomaly.
[0138] When the same current detection window simultaneously meets the judgment conditions for two or more types of anomalies, the window-level anomaly type result is determined in the order of heat recovery disorder anomaly, diffusion hindrance anomaly, heat source enhancement anomaly, and sensor drift anomaly.
[0139] Perform continuity filtering and stability filtering on window-level anomaly type results to generate anomaly type results;
[0140] The continuity screening and stability screening check whether the window-level anomaly type results of adjacent windows are consistent according to the current detection window order, and count the number of times the same anomaly type appears continuously in multiple consecutive current detection windows. When the number of times the same anomaly type appears continuously is not lower than the corresponding anomaly type duration threshold, the anomaly type is retained as an anomaly type result. When the same anomaly type does not reach the corresponding anomaly type duration threshold, and its intra-chain transport cost, inter-chain transport cost, heat recovery imbalance cost, new mismatch amount, and missing mismatch amount fall back to the corresponding low cost range in subsequent current detection windows, the window-level anomaly type result is deleted.
[0141] Based on the anomaly type results, an anomaly level determination is performed on each current detection window to generate an anomaly level result;
[0142] The anomaly level determination is based on the intra-chain transport cost, inter-chain transport cost, heat recovery imbalance cost, new mismatch amount, missing mismatch amount, and the number of times the anomaly type continuously occurs in the current detection window. These are compared with the first-level, second-level, third-level, and fourth-level ranges of the corresponding anomaly type to generate an anomaly level. When different indicators in the same current detection window correspond to different anomaly levels, the higher anomaly level is taken as the anomaly level result of the current detection window.
[0143] Extract the dominant anomaly region based on the anomaly type and anomaly level results;
[0144] The extraction of dominant anomaly regions involves sorting all node ranges corresponding to the node dominance distribution in each current detection window according to the node dominance magnitude, selecting the node ranges at the top of the sort and whose cumulative node dominance reaches a preset dominance ratio as candidate dominant regions. When candidate dominant regions are located within the same heat propagation chain, the continuous node ranges on the corresponding heat propagation chain are merged into intra-chain dominant anomaly regions. When candidate dominant regions belong to two or more heat propagation chains with inter-chain relationships, the corresponding node ranges are merged into inter-chain dominant anomaly regions. The anomaly type result, anomaly level result, and dominant anomaly region are then encapsulated according to the current detection window order.
[0145] Example 1: To verify the feasibility of this invention in practice, it was applied to a transmission temperature monitoring scenario in an automobile manufacturing company. The application location was a vehicle assembly workshop and a closed-road durability testing site in East China, covering high-temperature summer operation, urban congestion periods, and continuous high-speed driving. In this scenario, the transmission operates under alternating conditions of frequent start-stop cycles, continuous gear shifting, and high-load transmission. Localized heating of the housing, delayed oil return, heat transfer across regions, and localized drift of the temperature sensor are prone to occur. Existing methods mainly rely on single-point temperature thresholds, simple temperature rise amplitudes, or empirical correction rules to determine anomalies. When encountering situations such as heat transfer between different heat transfer paths, increased local heat sources without a significant overall temperature increase, or false anomalies caused by drift at a single measuring point, the conclusion is often that the temperature is too high. It is difficult to determine whether the anomaly originates from a genuine thermal fault or a data acquisition error, and it is also difficult to locate the main anomaly area.
[0146] In this scenario, temperature sensors are first deployed in the input housing, output housing, the area adjacent to the front bearing, the area adjacent to the rear bearing, the area adjacent to the oil return, the oil pan, and the side wall of the housing. Temperature sequences from multiple fixed measurement locations are continuously collected, and time alignment, abnormal jump point removal, and continuity verification are performed to form a set of temperature nodes. Subsequently, healthy temperature data is extracted from the long-term stable operation phase. The arrival order of temperature changes at each temperature node and the heat transfer correlation are compared to generate a thermal causality graph. Then, heat propagation baseline information within different detection windows is extracted from the healthy operation phase to form a baseline feature set. For the current temperature sequence collected during road verification and bench-based operation, time lag deviation, attenuation deviation, regeneration deviation, and node temperature difference deviation are extracted according to the same window rules to form a thermal mismatch feature set, which is then input into a non-equilibrium hierarchical Kantorovich transport model with thermal causality constraints. The model separates deviations within the same heat propagation chain from cross-chain transfers between different heat propagation chains. It determines whether local heat propagation is unbalanced in the intra-chain transport layer and whether the regeneration path is disordered in the inter-chain transport layer. It also retains newly added and missing mismatches to distinguish between heat source enhancement, diffusion hindrance, regeneration disorder, and sensor drift.
[0147] To ensure the model adapts to the actual thermal propagation characteristics of the transmission, a non-equilibrium hierarchical Kantorovich transport model with thermal causality constraints was trained during implementation. Training data consisted of temperature samples from the workshop's thermal break-in phase, road durability phase, and repair / disassembly records. Healthy operation data was used to generate a baseline feature set, while temperature samples from confirmed abnormal phases were used to construct a thermal mismatch feature set. These samples, combined with maintenance conclusions, were used to manually label anomaly types. During training, the permissible deviation ranges for propagation delay, attenuation relationship, and regenerative correlation, as well as various cost parameters in intra-chain and inter-chain transport, were continuously updated. This ensured that the model's output transport judgments were consistent with known anomaly types and that the distribution of new mismatches, missing mismatches, and node dominance remained stable across adjacent windows. After multiple rounds of training, the model was able to stably separate anomalies caused by enhanced local friction, diffusion hindrance caused by limited oil passages, regenerative turbulence caused by return oil path disturbances, and sensor drift caused by single-point temperature drift during the validation phase.
[0148] Data comparison during application shows that, after adopting this invention, the original method of judging solely by temperature level is replaced by a method based on the heat propagation structure. When replaying temperature data from the same time period, it can be observed that in scenarios with enhanced local heat sources, the intra-chain transport cost continuously increases and concentrates in the upstream node range; in scenarios with diffusion hindrance, the missing mismatch extends downstream along the heat transfer direction; in scenarios with chaotic regeneration, the inter-chain transport cost and the regeneration imbalance cost increase simultaneously; and in scenarios with sensor drift, the anomaly is mainly concentrated in a single node range and does not spread to adjacent heat propagation edges. This not only allows for earlier anomaly detection but also identifies the dominant anomaly area, facilitating rapid judgment by on-site personnel whether it is localized heating, heat transfer obstruction, regeneration imbalance, or an anomaly at the measuring point itself. Implementation results demonstrate that this invention effectively solves the problems of difficulty in distinguishing anomaly mechanisms, difficulty in locating anomaly areas, and easy misjudgment of single-point drift in existing technologies, possessing good feasibility and field application value.
[0149] Table 1. Comparison of Overall Performance in Transmission Temperature Anomaly Detection
[0150]
[0151] As shown in Table 1, although the over-temperature alarm algorithm based on a fixed threshold has a fast processing speed, its anomaly detection accuracy is only 74.6%, with a false alarm rate of 15.8% and a false negative rate of 17.9%. This indicates that this type of method mainly relies on whether the temperature at a single point exceeds the limit for judgment, making it difficult to distinguish between short-term thermal fluctuations, local heat accumulation, and sensor drift. Therefore, it is prone to misjudgment and false negative under complex operating conditions. The ARIMA temperature prediction residual alarm algorithm and the XGBoost temperature regression residual alarm algorithm can improve the detection results by utilizing time series change trends or nonlinear regression relationships. The anomaly detection accuracy is increased to 78.9% and 82.7%, respectively, and the false alarm rate is reduced to 12.7% and 10.4%, respectively. However, their judgment basis still mainly remains at the level of the deviation between the temperature value and the predicted value. They do not adequately characterize the coupling relationship between different heat transfer paths inside the gearbox. Therefore, the accuracy of anomaly type judgment and the accuracy of locating the dominant anomaly area are still relatively low.
[0152] The LSTM autoencoder reconstruction error detection algorithm and the Wasserstein distance change point detection algorithm have further improved overall performance, with anomaly detection accuracies reaching 84.1% and 85.3%, respectively, and average warning lead times reaching 12.4 minutes and 13.7 minutes, respectively. This indicates that these two algorithms are more adaptable to complex temporal fluctuations and distribution changes. In particular, the Wasserstein distance change point detection algorithm has outperformed the aforementioned regression-based methods in controlling false alarm and false negative rates, indicating that distribution-level comparison is more suitable than a single residual threshold for identifying slowly changing and abrupt thermal anomalies. However, these methods still mainly detect overall distribution shifts and do not incorporate the direction of heat propagation, propagation delay, attenuation relationship, and regeneration correlation within the gearbox into a unified constraint. Therefore, the accuracy in anomaly type determination and the accuracy in locating the dominant anomaly area are still only 72.1% and 65.4%, respectively, making it difficult to fully reflect the differences between local heat propagation hindrance, cross-chain regeneration disorder, and single-point drift.
[0153] The non-equilibrium hierarchical Kantorovich transport method with thermal causality constraints demonstrates superior overall performance across various metrics. Its anomaly detection accuracy reaches 91.8%, a 6.5 percentage point improvement over the Wasserstein distance change point detection algorithm. The false alarm rate decreases to 5.7%, and the false negative rate decreases to 7.4%, indicating that this method can more effectively distinguish between genuine thermal anomalies and short-term disturbances under complex operating conditions. Simultaneously, the anomaly type determination accuracy reaches 84.6%, and the dominant anomaly region location accuracy reaches 80.9%, significantly higher than other comparative algorithms. This suggests that this method not only improves the ability to determine whether an anomaly exists but also enhances the identification of anomaly mechanisms and locations. The average warning lead time reaches 18.6 minutes, an increase of 11.8 minutes compared to the fixed threshold overheat alarm algorithm and 4.9 minutes compared to the Wasserstein distance change point detection algorithm. This indicates that this method can detect potential risks earlier, before the overall temperature has significantly increased, through changes in thermal propagation structure deviation, cross-chain transfer, and regenerative imbalance. Although the single-window processing time is 53 milliseconds, which is higher than that of the threshold method and the regression method, it is still within the acceptable range for online monitoring and can meet the real-time requirements of workshop monitoring and road verification scenarios.
[0154] The aforementioned performance improvements primarily stem from three aspects. First, the temperature node set and the thermal cause-effect graph jointly introduce heat propagation direction, propagation time lag, attenuation relationship, and regeneration correlation, expanding the detection object from single-point temperature values to heat propagation structural features. Therefore, it can identify local heat source enhancement and heat propagation hindrance that are difficult to express explicitly using existing methods. Second, the non-equilibrium hierarchical Kantorovich transport model models deviations within the same heat propagation chain and cross-chain transfers between different heat propagation chains separately. It retains mismatch information through added and missing mismatch amounts, enabling the model not only to detect anomalies but also to reflect whether the anomaly originates from locally added heat, insufficient downstream heat transfer, or disordered regeneration links. Third, the node dominance distribution and regeneration imbalance cost provide direct evidence for anomaly region location. This allows the dominant anomaly region to be determined based on cost distribution in the transport results, rather than relying on empirical judgment, thus exhibiting a significant advantage in location accuracy.
[0155] In summary, this method, without relying on additional vibration, pressure, or speed measurement signals, achieves modeling of the heat propagation structure of the gearbox, hierarchical determination of complex temperature anomalies, and effective localization of dominant anomaly areas solely based on temperature sequences from multiple fixed temperature measurement locations. This improves the accuracy and early warning capabilities of anomaly detection, enhances the ability to distinguish anomaly types and the interpretability of results, while maintaining processing efficiency that meets the requirements of online applications. Therefore, it is more suitable for gearbox temperature monitoring and anomaly diagnosis under complex operating conditions.
[0156] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for detecting gearbox temperature based on a temperature sensor, characterized in that, Includes the following steps: Collect temperature sequences from various fixed temperature measurement locations in the gearbox and construct a set of temperature nodes; A heat cause-effect graph is generated based on the sequential response relationship and heat transfer correlation of each node in the temperature node set. The detection windows are divided according to the continuous sampling period. The heat propagation baseline information of each detection window in the healthy operation phase is extracted based on the heat causality diagram to generate a baseline feature group. Based on the temperature sequence and thermal cause-effect graph of the current detection window, thermal mismatch information characterizing the current thermal propagation deviation state is extracted, and thermal mismatch feature groups are generated. Input the baseline feature set and the thermal mismatch feature set into the non-equilibrium hierarchical Kantorovich transport model with thermal causality constraints to generate transport determination results; Based on the transport determination results, determine whether the anomaly is heat source enhancement type, diffusion hindrance type, regenerative turbulence type, or sensor drift type, and output the anomaly level and dominant anomaly area.
2. The method for detecting gearbox temperature based on a temperature sensor according to claim 1, characterized in that, The generation of the temperature node set specifically includes: Fixed temperature measurement positions are set in the input housing, output housing, front bearing adjacent area, rear bearing adjacent area, oil return adjacent area, oil pan area and housing side wall area of the gearbox, each corresponding to a unique node identifier, forming a candidate temperature node list; The sampled temperature at each fixed temperature measurement location in the candidate temperature node list is continuously sampled according to a uniform sampling period, and the sampling results of each fixed temperature measurement location are sorted in chronological order to form the original temperature sequence group. Time alignment, outlier removal, and continuity verification are performed on the original temperature sequence group to retain valid temperature segments that meet the requirements of a unified time base and continuous sampling, thus forming a regular temperature sequence group. Each effective temperature segment in the regular temperature sequence group is bound to its corresponding node identifier. Each node identifier corresponds to a unique regular temperature sequence, forming a temperature node. Temperature nodes are selected and encapsulated based on their fixed positions on the gearbox and the continuous effectiveness of the regular temperature sequence, thus generating a set of temperature nodes.
3. The method for detecting gearbox temperature based on a temperature sensor according to claim 1, characterized in that, The generation of the heat cause-effect graph specifically includes: Taking the regular temperature sequence corresponding to each temperature node in the temperature node set as the object, the arrival order of temperature changes of each temperature node is compared one by one under a unified time reference. The upstream and downstream nodes and their corresponding sequential relationships of each node pair in each comparison period are recorded to form a node sequential response record set. Perform stability screening on the node response record set, remove node pairs whose arrival order is repeatedly reversed, and retain node pairs whose arrival order remains consistent over multiple consecutive comparison periods to form a candidate response relationship set; A heat transfer correlation record set is formed by checking the continuous following state, stable decay state and heat recovery corresponding state between the nodes in the candidate response relationship set. For candidate heat transfer node pairs in the heat transfer association record set, extract the propagation time delay base value, attenuation relationship and heat recovery association to form edge attribute groups, and bind them with the corresponding candidate heat transfer nodes to form heat propagation edges; According to the connection direction of the upstream and downstream nodes of each heat propagation edge, heat propagation edges with continuous transmission relationship are sequentially spliced together to form a set of heat propagation chains; Using the set of temperature nodes as graph nodes, the set of heat propagation edges as graph edges, and the set of heat propagation chains as link constraints, the propagation time delay base value, attenuation relationship and heat recovery association in the edge attribute group are written into the corresponding graph edges to generate a thermal cause-effect graph.
4. The method for detecting gearbox temperature based on a temperature sensor according to claim 3, characterized in that, The edge attribute group is obtained by summarizing the delay duration, attenuation ratio and heat recovery correspondence ratio of candidate heat transfer node pairs within the effective comparison period. For each candidate heat transfer node pair, the comparison period that simultaneously satisfies the conditions of stable order, continuous following, stable attenuation and heat recovery correspondence is retained as the effective comparison period. The propagation time delay base value, attenuation relationship and heat recovery association are calculated respectively. The same candidate heat transfer node pair and its corresponding propagation time delay base value, attenuation relationship and heat recovery association are encapsulated as heat propagation edge.
5. The method for detecting gearbox temperature based on a temperature sensor according to claim 1, characterized in that, The generation of the baseline feature set specifically includes: In the regular temperature sequence corresponding to the set of temperature nodes, a continuous sampling period with continuous and effective temperature nodes, stable node response relationship and stable heat transfer correlation is selected as the healthy operation stage. The healthy operation stage is divided according to a unified window duration and a unified sliding step size to form a health detection window sequence. Around each health detection window, based on the set of heat propagation chains in the heat propagation graph, the connection direction of heat propagation edges, and the base value of propagation time delay, attenuation relationship and heat recovery association of each heat propagation edge, the link order and corresponding node range of each heat propagation chain within the health detection window are determined, forming a link benchmark mapping table. According to the link reference mapping table, extract the heat propagation reference quantity of each heat propagation chain in each health detection window to form a window-level heat propagation reference segment. Window-level heat propagation reference segments belonging to the same heat propagation chain are merged according to the heat propagation chain identifier. Consistency screening and central summarization are performed on the merged window-level heat propagation reference segments to form chain-level reference entries. The chain-level benchmark entries corresponding to each heat propagation chain are encapsulated according to the order of the heat propagation chain set to generate benchmark feature groups.
6. The method for detecting gearbox temperature based on a temperature sensor according to claim 1, characterized in that, The generation of the thermal mismatch feature set specifically includes: The current continuous sampling period is divided into the current detection window sequence according to the same window duration and sliding step size as the health detection window sequence. Based on the set of heat propagation chains, the connection direction of heat propagation edges, and the base value of propagation time delay, attenuation relationship and heat recovery association of each heat propagation edge in the heat cause-effect graph, the link order and corresponding node range of each heat propagation chain in each current detection window sequence are determined to form the current link mapping table. According to the current link mapping table, extract the current heat propagation amount of each heat propagation chain in each current detection window to form a window-level current heat propagation segment; The window-level current heat propagation segment is aligned with the corresponding chain-level benchmark entry in the benchmark feature group according to the heat propagation chain identifier, heat propagation edge order and node range. The time delay deviation result, attenuation deviation result, heat recovery deviation result and node temperature difference deviation result are extracted to form a window-level heat mismatch segment. The window-level thermal mismatch segments corresponding to each heat propagation chain within the same current detection window are merged and sequentially encapsulated to form the thermal mismatch information corresponding to the current detection window. The thermal mismatch information corresponding to each current detection window in the current detection window sequence is encapsulated according to the order of the detection windows to generate thermal mismatch feature groups.
7. The method for detecting gearbox temperature based on a temperature sensor according to claim 1, characterized in that, The generation of the transportation determination result specifically includes: A non-equilibrium hierarchical Kantorovich transport model with thermal causality constraints is constructed around the baseline feature set, the thermal mismatch feature set, and the thermal causality graph. According to the heat propagation chain set, the baseline feature group and the thermal mismatch feature group are paired and split in the same chain, and the pairing results are written into the intra-chain transport layer to form an intra-chain transport unit set. Inter-chain relationships are constructed around the set of intra-chain transport units and written into the inter-chain transport layer to form a hierarchical transport structure; Constraints are configured on the hierarchical transport structure in thermal causality constraint units and non-equilibrium constraint units to form a constraint transport cost table; In the intra-chain transport layer, position matching and cost accumulation are performed according to the constraint transport cost table to obtain the intra-chain transport result; In the inter-chain transport layer, cross-chain transfers and reheat checks are performed based on the constraint transport cost table to obtain the inter-chain transport results. According to the current detection window order, the intra-chain transport results and inter-chain transport results corresponding to each heat propagation chain are merged to generate transport judgment results.
8. The method for detecting gearbox temperature based on a temperature sensor according to claim 7, characterized in that, The non-equilibrium hierarchical Kantorovich transport model with thermal causal constraints uses a reference feature set as the reference side input, a thermal mismatch feature set as the mismatch side input, and a thermal causal graph as the link constraint input. It includes an intra-link transport layer, an inter-link transport layer, thermal causal constraint units, and non-equilibrium constraint units. The intra-chain transport layer is used to perform position matching and cost accumulation between the reference side and the mismatch side within the same heat propagation chain. The inter-chain transport layer is used to perform cross-chain matching and cost accumulation for mismatch transfer between different heat propagation chains. The thermal causality constraint unit is used to constrain the transport direction, propagation time delay, attenuation relationship and regeneration correlation. The non-equilibrium constraint unit is used to retain and accumulate the mismatch and missing parts that cannot be matched.
9. The method for detecting gearbox temperature based on a temperature sensor according to claim 1, characterized in that, The generation of the anomaly level and the dominant anomaly region specifically includes: Extract the transport judgment results corresponding to each current detection window in the current detection window order, and extract the intra-chain transport cost, inter-chain transport cost, new mismatch, missing mismatch, node dominance distribution and heat recovery imbalance cost for each transport judgment result to form an anomaly judgment sequence. Based on the anomaly determination sequence, anomaly type determination is performed on each current detection window to form a window-level anomaly type result; Perform continuity filtering and stability filtering on window-level anomaly type results to generate anomaly type results; Based on the anomaly type results, an anomaly level determination is performed on each current detection window to generate an anomaly level result; Extract the dominant anomaly region based on the anomaly type and anomaly level results.
