A crane gear box fault diagnosis method, electronic equipment and storage medium

Abstract

This application provides a method, electronic device, and storage medium for diagnosing crane gearbox faults. The method includes: dividing a specified time period into at least one continuous actual time-series segment; determining the type of preset operating condition for each actual time-series segment; obtaining the operating condition parameters of the crane gearbox in a healthy state for each actual time-series segment; obtaining predicted vibration time-series data for each specified location; obtaining a first physical residual sequence between the predicted vibration time-series data and the actual vibration time-series data; obtaining a first initial fault diagnosis result for the crane gearbox in each actual time-series segment using a deep learning model; verifying the first initial fault diagnosis result based on the actual temperature and actual rotational speed at each time step in the actual time-series segment to obtain a second initial fault diagnosis result; and obtaining a final diagnosis result for the crane gearbox. This solution ensures the reliability of the final diagnosis result.

Description

Technical Field

[0001] This invention relates to the field of special equipment fault diagnosis technology, and more specifically, to a method for diagnosing faults in a crane gearbox, an electronic device, and a storage medium. Background Technology

[0002] Cranes are core heavy-duty equipment in industrial production and infrastructure projects. As a key component of their power transmission, the gearbox is subjected to variable loads, variable speeds, impact loads, and reciprocating alternating stresses over long periods, making it highly susceptible to faults such as tooth surface wear, pitting, tooth breakage, bearing spalling, gear eccentricity, and shaft loosening. Gearbox faults are characterized by their high degree of concealment, weak early signs, rapid evolution, and high cascading risks. A sudden failure can directly lead to crane downtime, operational interruption, and even major safety accidents such as equipment overturning and personal injury. Therefore, accurate fault diagnosis of crane gearboxes is of significant engineering importance for safe equipment operation and maintenance, early warning of potential hazards, and reduction of operation and maintenance costs.

[0003] Current mainstream methods for fault diagnosis of crane gearboxes primarily rely on data-driven deep learning approaches. These methods collect gearbox vibration signals and use models such as convolutional neural networks and recurrent neural networks to classify and identify faults. However, existing crane gearbox fault diagnosis methods do not differentiate between operating conditions for continuously collected vibration signals. They use a uniform model and threshold to process data from all operating conditions and rely solely on vibration signals as the sole diagnostic criterion. They fail to incorporate the gearbox's dynamic transmission mechanism to construct physical expectations under healthy conditions, nor do they utilize independent physical quantities such as temperature and rotational speed to cross-check the diagnostic results. This leads to inconsistent diagnostic standards under different operating conditions, early fault characteristics being submerged by background noise and difficult to separate, and non-fault vibration interference being easily misidentified as internal faults. Overall, the diagnostic accuracy is low, the false alarm rate is high, and the results lack physical basis. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a crane gearbox fault diagnosis method, electronic device and storage medium, which can improve the problems of low overall diagnostic accuracy, high false alarm rate and lack of physical basis of the results in the prior art.

[0005] To achieve the above technical objectives, the technical solution adopted in this application is as follows:

[0006] In a first aspect, embodiments of this application provide a method for diagnosing faults in a crane gearbox, the method comprising:

[0007] The actual rotational speed, actual temperature, and actual load of the crane gearbox at each time step within a specified period are obtained, as well as the actual vibration parameters of the crane gearbox at at least one specified location, to obtain actual rotational speed time-series data, actual temperature time-series data, actual load time-series data, and actual vibration time-series data at the specified location.

[0008] Based on the actual rotational speed, actual temperature, and actual load of each time step, the specified time period is divided into at least one continuous actual time sequence segment;

[0009] Based on the actual rotational speed, actual temperature, and actual load of each time step in the actual time sequence segment, the type of preset operating condition for each actual time sequence segment is determined, and the number of preset operating conditions is several.

[0010] In a pre-established health baseline library, the operating parameters of each actual time segment of the crane gearbox in a healthy state are obtained;

[0011] The actual rotational speed time series data, actual load time series data, and operating condition parameters of each actual time series segment are input into the gearbox model. The gearbox model generates the predicted vibration parameters of the specified position at each time step through forward simulation, so as to obtain the predicted vibration time series data of each specified position.

[0012] The first physical residual sequence between the predicted vibration time series parameters and the actual vibration time series parameters at each time step is obtained by using the difference between the predicted vibration time series data and the actual vibration time series data.

[0013] The first physical residual sequence, the actual vibration time series data, and the identifier of the actual time series segment are input into the deep learning model. The deep learning model is used to obtain the first initial fault diagnosis result of the crane gearbox when it is in each actual time series segment. The first initial fault diagnosis result includes whether the crane gearbox is faultless when it is in the corresponding actual time series segment, or whether there is a fault of a specified fault type.

[0014] Based on the actual temperature and actual rotation speed of each time step in the actual time sequence segment, the first initial fault diagnosis result is verified to obtain a second initial fault diagnosis result. The second initial fault diagnosis result indicates whether the first initial fault diagnosis result is credible or unreliable.

[0015] The final diagnosis result of the crane gearbox is obtained based on all the second initial fault diagnosis results that characterize the credibility of the first initial fault diagnosis result.

[0016] 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 described in the first aspect.

[0017] Thirdly, embodiments of this application also provide a computer-readable storage medium storing a computer program that, when run on a computer, causes the computer to perform the method described in the first aspect.

[0018] The invention employing the above technical solution has the following advantages:

[0019] In the technical solution provided in this application, multi-source synchronous time-series data is formed by acquiring the actual rotational speed, actual temperature, actual load, and actual vibration parameters at at least one specified location of the crane gearbox at each time step within a specified period, providing comprehensive operating status information for subsequent analysis. By dividing the specified period into continuous actual time-series segments based on the actual rotational speed, actual temperature, and actual load at each time step, and determining the preset operating condition type for each segment, data under different operating conditions are processed separately, avoiding the problem of inconsistent diagnostic standards caused by mixed analysis of different operating conditions. The operating condition parameters corresponding to each actual time-series segment are obtained from a pre-established health baseline library, and the actual rotational speed time-series data, actual load time-series data, and operating condition parameters are input into the gearbox model for forward simulation to generate predicted vibration time-series data, providing a physical mechanism-based healthy vibration expectation for each operating condition. The difference between the predicted vibration and the actual vibration is calculated. The first physical residual sequence is obtained by removing the expected vibration components determined by rotational speed and load under healthy conditions, making the fault characteristics relatively enhanced in the residual, which facilitates the identification of early and minor faults. By inputting the first physical residual sequence, actual vibration time series data, and actual time series segment identifiers into the deep learning model, the first initial fault diagnosis result is obtained, which integrates physical residual features, original vibration features, and operating condition context features, thereby improving the accuracy of fault identification. The second initial fault diagnosis result is obtained by verifying the first initial fault diagnosis result based on actual temperature and actual rotational speed. The vibration diagnosis conclusion is cross-validated using the two independent physical quantities of temperature and rotational speed, which suppresses false alarms caused by sensor interference and external transient impacts. The final diagnosis result is obtained by using all the second initial fault diagnosis results that are deemed credible, and only the diagnosis conclusion verified by temperature and rotational speed is accepted, further ensuring the reliability of the final diagnosis result. Attached Figure Description

[0020] 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.

[0021] Figure 1 A flowchart of a crane gearbox fault diagnosis method provided in an embodiment of this application. Detailed Implementation

[0022] 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.

[0023] Please refer to Figure 1 This application provides a method for diagnosing faults in a crane gearbox, which can be applied to electronic devices and whose steps can be executed or implemented by the electronic devices. The electronic devices can be, but are not limited to, personal computers, smartphones, etc. The crane gearbox fault diagnosis method may include the following steps:

[0024] S110, acquire the actual rotational speed, actual temperature and actual load of the crane gearbox at each time step in a specified period, as well as the actual vibration parameters of the crane gearbox at at least one specified position, to obtain actual rotational speed time-series data, actual temperature time-series data, actual load time-series data and actual vibration time-series data at the specified position;

[0025] S120, based on the actual rotational speed, actual temperature and actual load of each time step, the specified time period is divided into at least one continuous actual time sequence segment;

[0026] S130, based on the actual rotational speed, actual temperature and actual load of each time step in the actual time segment, determine the type of preset working condition for each actual time segment, and the number of preset working conditions is several;

[0027] S140, Obtain the operating parameters of each actual time segment of the crane gearbox in a healthy state from the pre-established health baseline library;

[0028] S150, the actual rotational speed time series data, actual load time series data, and working condition parameters of each actual time series segment are input into the gearbox model. The gearbox model generates the predicted vibration parameters of the specified position at each time step through forward simulation, so as to obtain the predicted vibration time series data of each specified position.

[0029] S160, by using the difference between the predicted vibration time series parameters and the actual vibration time series parameters at each time step, the first physical residual sequence between the predicted vibration time series data and the actual vibration time series data is obtained;

[0030] S170, the first physical residual sequence, the actual vibration time series data and the identifier of the actual time series segment are input into the deep learning model, and the first initial fault diagnosis result of the crane gearbox when it is in each actual time series segment is obtained through the deep learning model. The first initial fault diagnosis result includes no fault in the crane gearbox when it is in the corresponding actual time series segment, or a fault of a specified fault type exists.

[0031] S180, based on the actual temperature and actual rotation speed of each time step in the actual time segment, verify the first initial fault diagnosis result to obtain a second initial fault diagnosis result, wherein the second initial fault diagnosis result indicates whether the first initial fault diagnosis result is credible or unreliable;

[0032] S190, based on all the second initial fault diagnosis results that characterize the credibility of the first initial fault diagnosis result, the final diagnosis result of the crane gearbox is obtained.

[0033] This technical solution divides a specified time period into continuous actual time-series segments and determines the preset operating condition type for each segment. The gearbox model then uses the corresponding operating condition parameters from the health baseline library to perform forward simulations for different operating conditions, thus establishing a unique healthy vibration expectation for each condition. This avoids the diagnostic standard confusion caused by using a unified model and threshold to process all operating condition data. The gearbox model generates predicted vibration time-series data based on operating condition parameters, actual speed, and actual load. The predicted vibration time-series data is then subtracted point-by-point from the actual vibration time-series data to obtain the first physical residual sequence. This allows background vibration components under healthy conditions to be canceled out in the difference, while abnormal vibration components caused by faults are relatively amplified in the residual. This separates early, weak fault characteristics from background noise, providing physically meaningful data for the deep learning model. Fault enhancement features: By inputting the first physical residual sequence, actual vibration time series data, and the identifier of the actual time series segment into the deep learning model, the first initial fault diagnosis result is obtained. This allows the model to obtain the assistance of physical residual features in addition to the original vibration features, and the diagnosis result has a dynamic mechanism basis. Furthermore, the first initial fault diagnosis result is verified by using the actual temperature and actual rotational speed of each time step in the actual time series segment. The vibration anomaly is verified to have a real physical root by using two physical quantities independent of vibration, namely temperature and rotational speed. Vibration anomalies that lack corresponding changes in temperature and rotational speed are judged as unreliable. The final diagnosis result is obtained only based on the second initial fault diagnosis result with reliable characterization. This eliminates false alarms caused by sensor interference or external transient impacts, thereby improving the overall diagnostic accuracy and reducing the false alarm rate.

[0034] The following is a detailed explanation of each step in the crane gearbox fault diagnosis method:

[0035] The parameters in S110 can be obtained in the following way:

[0036] Temperature sensors are installed on the outer wall of the bearing end cover, the oil sump wall, or the surface of the gearbox housing of the crane gearbox. These sensors are either platinum resistance temperature sensors or thermocouple sensors, fixed to the surface being measured by threaded installation or thermally conductive adhesive, ensuring close contact between the sensor and the metal surface to minimize temperature measurement hysteresis. The output signal of the temperature sensor is converted into a standard voltage or current signal by a temperature transmitter module and then connected to the temperature channel of the data acquisition system. The data acquisition system continuously samples the temperature signal at a preset sampling frequency. Since temperature is a slowly changing physical quantity, the temperature sampling frequency can be lower than the vibration sampling frequency (e.g., once per second to ten times per second), but it must be ensured that the temperature data and vibration data are aligned on the same time axis. Each sampling point corresponds to a timestamp, and all sampling points are arranged in chronological order to obtain the actual temperature time-series data. If the gearbox's existing lubrication system or monitoring system is already equipped with temperature monitoring points, the temperature data can be directly read from the existing system and, after timestamp alignment, used as the actual temperature time-series data.

[0037] Load sensors are installed on the fixed pulley blocks, movable pulley blocks, or fixed ends of the wire rope in the crane's hoisting mechanism. These load sensors are any one of pin-type load cells, plate-ring type tension sensors, or pressure-side load cells, connected in series in the wire rope's force path to detect the weight of the hoisting load in real time. The output signal of the load sensor is amplified by a weighing transmitter module and converted into a standard voltage or current signal before being connected to the load channel of the data acquisition system. The data acquisition system continuously samples the load signal at a preset sampling frequency, which is the same as or close to the rotational speed sampling frequency to ensure a complete record of the load change process. Each sampling point corresponds to a timestamp, and all sampling points are arranged in chronological order to obtain the actual load time-series data. If the crane is equipped with a lifting weight limiter or torque limiter, the load data can be directly read from the communication interface of this device and, after timestamp alignment, used as the actual load time-series data. For gearboxes that do not directly participate in hoisting (such as luffing mechanisms or slewing mechanism gearboxes), a torque sensor is installed on the transmission path of the corresponding mechanism, or the output torque is read from the motor drive system as the equivalent load.

[0038] A vibration acceleration sensor is installed at a designated location on the crane gearbox housing. This designated location includes at least one of the following: the end face of the input shaft bearing housing, the end face of the output shaft bearing housing, and the housing wall corresponding to the gear meshing area. The vibration acceleration sensor is rigidly fixed to the surface being measured using a magnetic mount, threaded installation, or epoxy adhesive bonding, ensuring no relative looseness between the sensor and the housing to avoid introducing additional vibration. The output signal of the vibration acceleration sensor is amplified by a charge amplifier or a built-in integrated circuit and then input to the vibration channel of the data acquisition system. The data acquisition system continuously and synchronously acquires the vibration signal at a high sampling frequency of 20 kHz to 50 kHz, with a sampling frequency no less than five times the highest meshing frequency of the gearbox, to ensure complete capture of gear meshing impact and bearing fault characteristic frequencies. Each sampling point corresponds to a timestamp, and all sampling points are arranged in chronological order to obtain the actual vibration time-series data for the designated location. When there are multiple designated locations, each location corresponds to one vibration channel. All vibration channels are synchronously sampled in hardware within the data acquisition system to ensure strict consistency of timestamps between vibration signals.

[0039] In S120, this is specifically achieved in the following way:

[0040] Based on the actual rotational speed, actual temperature, and actual load of the time step, and according to a preset rule, a time step for operating condition switching is determined among all the time steps; based on the time step for operating condition switching, the actual time sequence segment is determined; the preset rule indicates that if the rate of change of the actual rotational speed corresponding to a first specified number of consecutive time steps exceeds a first threshold, then the last time step of the specified number of time steps is the time step for operating condition switching; the preset rule indicates that if the absolute value of the rate of change of the actual load corresponding to a second specified number of consecutive time steps exceeds a second threshold, then the last time step of the specified number of time steps is the time step for operating condition switching; the preset rule indicates that when the direction of the actual rotational speed changes abruptly, the time step of the abrupt change in direction is the time step for operating condition switching.

[0041] First, the actual speed change rate and actual load change rate are pre-calculated for all time steps within a specified time period. For each time step, linear fitting is performed on the speed and load values ​​of several adjacent time steps centered on itself to obtain the absolute values ​​of the speed change rate and load change rate for that time step, which are stored with each time step. Then, starting from the first time step, the process is traversed sequentially, and three preset rules are used in parallel and independently to detect candidate switching points: Rule 1 maintains a continuous counter. When the speed change rate of a first specified number of consecutive time steps exceeds the first threshold, the last time step of that segment is recorded as a candidate switching point and the counter is reset to zero, used to identify switching moments of continuous acceleration and deceleration. Similarly, Rule 2 maintains a continuous counter. When the absolute value of the load change rate of a second specified number of consecutive time steps exceeds the second threshold, the last time step of that segment is recorded as a candidate switching point, used to identify moments of sudden load changes such as cargo lifting off or landing. Rule 3 compares the speed direction of the current time step with that of the previous time step. When there is a direct jump between forward and reverse rotation, the current time step is recorded as a candidate switching point, used to identify the reverse switching moments of lifting and lowering. Each of the three rules outputs a set of candidate switching points without interfering with each other.

[0042] Candidate switching points according to the three rules are merged into a total set. First, duplicates are removed when multiple rules mark the same time step. Then, adjacent switching points with time intervals less than the minimum segment length are merged, and then sorted in ascending order of time to form an ordered switching point list. Using this list as a boundary, the first segment is defined as the period from the start of the specified time period to the first switching point, and the segments between adjacent switching points are defined as the subsequent segments. The segment from the last switching point to the end of the time period is defined as the last segment. Each segment carries start and end timestamps and all time series data. Finally, quality checks are performed on all segments. Segments with a duration less than the minimum diagnostic length are merged with the previous segment. Incomplete segments that are in the process of switching at the start or end of the time period are marked and their confidence is reduced. Empty segments are directly discarded. Finally, several continuous actual time series segments with qualified durations and clear boundaries are obtained.

[0043] This method identifies the moment of operating condition switching by detecting abrupt changes in three dimensions: actual speed change rate, absolute value of actual load change rate, and actual speed direction. The mechanism is as follows: the operating condition switching of the crane gearbox is essentially a step change in the equipment's operating state. Any operating condition switching is inevitably accompanied by a significant abrupt change in at least one physical quantity, such as speed, load, or rotation direction. Specifically, the speed change rate continuously exceeding the first threshold corresponds to the transition process of the equipment from uniform speed to acceleration / deceleration or from acceleration / deceleration to uniform speed, indicating a change in the dynamic response state of the transmission system. The absolute value of the load change rate continuously exceeding the second threshold corresponds to the process of goods leaving the ground, landing, or a significant adjustment of the load, indicating a fundamental change in the stress state and load level borne by the gearbox. The abrupt change in speed direction corresponds to the reverse switching between lifting and lowering, indicating a reversal of the power flow direction and meshing force direction of the gearbox. The three rules cover all possible operating condition switching types during crane operation from three complementary dimensions: dynamic response, load level, and power flow. This ensures that any real operating condition switching can be captured by at least one rule. The continuous data is divided into several actual time-series segments based on the detected switching time step. This ensures that the rotational speed, load, and rotation direction within each segment are in a relatively stable state, and the vibration characteristics within the segment have a consistent physical background and statistical properties. This provides a uniformly operating analysis unit for subsequent segment-based determination of preset operating condition types, invocation of corresponding healthy baseline parameters, generation of predicted vibration, and calculation of physical residuals. This avoids the problems of baseline parameter mismatch and residual calculation distortion caused by mixing data from different operating conditions in the same segment.

[0044] S130 specifically includes the following solutions:

[0045] S131: Based on the actual temperature of each time step in the actual time segment, obtain the actual temperature characteristics of each time step in the actual time segment, wherein the actual temperature characteristics include the actual temperature change rate;

[0046] S132: Based on the actual rotational speed of each time step in the actual time segment, obtain the actual rotational speed characteristics of each time step in the actual time segment, wherein the actual rotational speed characteristics include the actual rotational speed change rate, the actual rotational speed direction, and the actual rotational speed absolute value;

[0047] S133: Based on the actual load of each time step in the actual time segment, obtain the actual load characteristics of each time step in the actual time segment, wherein the actual load characteristics include the absolute value of the actual load;

[0048] S134: Based on the actual temperature characteristics, actual speed characteristics, and actual load characteristics of each time step in the actual time sequence segment, and based on the preset operating condition mapping table, determine the preset operating condition corresponding to each time step in the actual time sequence segment. The preset operating condition mapping table includes the mapping relationship between the preset operating condition and the first range of the actual temperature characteristics, the second range of the actual speed characteristics, and the third range of the actual load characteristics.

[0049] S135: When the number of time steps of the same preset working condition is greater than or equal to the third threshold, the corresponding preset working condition is the preset working condition of the actual time segment. When the number of time steps of the same preset working condition is less than the third threshold, the preset working condition of the actual time segment is determined according to the preset working conditions of the adjacent actual time segments.

[0050] In S131-S133, for each actual time segment, starting from its first time step and sequentially traversing to the last time step, three types of feature extraction operations are performed synchronously for each time step: For temperature feature extraction, taking the current time step as the center, the actual temperature values ​​of several time steps before and after it (usually taking data points corresponding to five seconds before and after) are linearly fitted. The slope obtained from the fitting is taken as the actual temperature change rate of that time step. Since temperature is a slowly varying physical quantity, the fitting neighborhood length is larger than the fitting neighborhood of rotational speed and load to suppress high-frequency noise. The actual temperature change rate characterizes the trend and rate of temperature rise and fall at that moment. For rotational speed feature extraction, firstly, the absolute value of the actual rotational speed value of the current time step is taken as the actual rotational speed absolute value. Secondly, the sign of the actual rotational speed value is taken as the actual rotational speed direction: positive values ​​are marked as forward rotation, negative values ​​as reverse rotation, and when the absolute value is below the direction determination dead zone, it is marked as stationary. Finally, taking the current time step as the center, the actual temperature values ​​of several time steps before and after it are linearly fitted. Linear fitting is performed on the actual rotational speed values ​​of several time steps (usually data points corresponding to 0.05 seconds before and after each step). The slope of the fitting is taken as the actual rotational speed change rate of that time step, which represents the rising and falling trend and acceleration of the rotational speed at that moment. In terms of load feature extraction, the actual load value of the current time step is directly taken as the absolute value of the actual load. If the current time step is at a moment of load change (i.e., the load change rate of its immediate vicinity exceeds the change threshold), it is additionally marked as a load transient point. The load transient point reduces the matching weight of the load feature during subsequent working condition matching. The extraction operations of the three features are completed sequentially on the same time step. The actual temperature change rate, actual rotational speed change rate, actual rotational speed direction, actual rotational speed absolute value, and actual load absolute value of each time step constitute the feature vector of that time step. The feature vectors of all time steps are arranged in chronological order to obtain the time-step feature sequence of the actual time segment, which is used for subsequent working condition type determination based on the preset working condition mapping table.

[0051] S134 specifically includes the following schemes:

[0052] Based on the mapping relationship recorded in the preset operating condition mapping table, the matching degree between the time step and the type of each preset operating condition is obtained through the matching score of the actual temperature feature with the first range, the matching score of the actual speed feature with the second range, and the matching score of the actual load feature with the third range, as well as the pre-calibrated fixed weights of the actual temperature feature, the actual speed feature, and the actual load feature; when the highest matching degree is greater than or equal to the fourth threshold and the highest matching degree is unique, the type of the highest preset operating condition is taken as the preset operating condition corresponding to each time step.

[0053] For each time step, the extracted actual temperature characteristics, actual speed characteristics, and actual load characteristics are compared with the first, second, and third ranges corresponding to each preset operating condition in the preset operating condition mapping table to calculate the matching degree: For temperature characteristic matching, it is checked whether the actual temperature change rate of the time step falls within the first range of the current preset operating condition; if it falls within, a full score is obtained for the temperature characteristic; otherwise, a score of zero is obtained. The first range is the normal range of temperature change rate under the healthy state of this operating condition. For speed characteristic matching, it is checked whether the absolute value of the actual speed falls within the sub-range of the absolute value of the speed in the second range, and whether ... The speed matching score is determined by three criteria: whether the actual rotational speed direction is consistent with the rotational speed direction in the second range, whether the actual rotational speed change rate falls within the sub-interval of the rotational speed change rate in the second range, whether all three criteria are met, whether the speed matching score is full, whether two criteria are met, whether one criterion is met, and whether none criteria are met, resulting in zero points. If the time step is a load transient point, the speed matching score is unaffected. In terms of load matching, the score is determined by checking whether the absolute value of the actual load falls within the third range. If it does, the load matching score is full; otherwise, zero points are awarded. If the time step is a load transient point, the load matching score is halved to reduce the impact of unreliable load data at transient moments on the judgment result. The matching scores of the three types of features are multiplied by pre-calibrated fixed weights and then summed to obtain the matching degree between the time step and the current preset working condition type. Among them, the speed feature has the highest weight (because speed is the most important basis for distinguishing the working condition type, and its absolute value, direction and rate of change can distinguish the differences between core working conditions such as acceleration, constant speed, braking, lifting and descent). The load feature has the second highest weight, and the temperature feature has the lowest weight (because temperature is a slowly changing signal and only plays an auxiliary verification role). The sum of the three weights is one. The weight values ​​are determined by optimizing the classification accuracy of health data during the system calibration stage. After traversing all preset operating condition types, the matching degree value between the time step and each operating condition type is obtained. The maximum value is taken. When the maximum matching degree is greater than or equal to the fourth threshold and the maximum value is unique among all operating condition types, the corresponding preset operating condition type is determined as the preset operating condition for the time step. When the maximum matching degree is lower than the fourth threshold, it means that the characteristics of the time step do not meet the typical characteristics of any preset operating condition. It is usually in the transient state of operating condition switching or at a time of data abnormality, and is marked as an uncertain point of operating condition. When the matching degree of two or more operating condition types is tied for the highest, it means that the characteristics of the time step meet the boundary conditions of multiple operating conditions at the same time. The operating condition is unclear, and is also marked as an uncertain point of operating condition.The mechanism is as follows: different preset working conditions have distinguishable feature ranges in the three physical dimensions of temperature, speed, and load. The degree of conformity between the features of a time step and the typical features of each working condition is measured by weighted matching. The speed has the highest weight to ensure that the working condition judgment is based on the motion state as the main basis. The load and temperature have lower weights as auxiliary verification. The fourth threshold and uniqueness requirement constitute a dual judgment threshold of absolute qualification and relative uniqueness. This avoids forcibly classifying transient points where features deviate from all working conditions, and also prevents misjudgment caused by feature ambiguity under various working condition boundary conditions. Thus, it ensures that the working condition judgment result of each time step has a clear physical basis and sufficient confidence.

[0054] In S135, when the number of time steps with the same preset operating condition is less than the third threshold, the type of preset operating condition is determined based on the preset operating conditions of adjacent actual time segments. This is achieved through the following scheme:

[0055] When adjacent actual time segments exist at both ends of the actual time segment in the time sequence, a second physical residual sequence of the actual time segment and the two adjacent actual time segments is calculated. The second physical residual sequence is obtained by the difference between the actual vibration parameter of the time step of the actual time segment and the actual vibration parameter of the time step of the adjacent actual time segment at the same position in the time sequence. The type of the preset operating condition of the adjacent actual time segment corresponding to the smaller value of the two second physical residual sequences is the type of the preset operating condition of the actual time segment. When adjacent actual time segments exist only at one end of the actual time segment in the time sequence, the type of the preset operating condition of the corresponding adjacent actual time segment is the type of the preset operating condition of the actual time segment.

[0056] When the number of time steps corresponding to all preset operating condition types within an actual time segment is less than the third threshold, the segment is determined to be a transitional segment with an unclear operating condition classification. In this case, its preset operating condition type is determined by comparing its vibration similarity with adjacent actual time segments. The specific implementation process is as follows: First, search the time segment list for adjacent actual time segments at the preceding and following ends of the transitional segment. If the adjacent segments are also transitional segments, continue searching in the corresponding direction until a stable segment with a clearly defined preset operating condition type is found or the specified time period boundary is reached. When there are valid adjacent actual time segments at both ends, perform two second physical residual sequence calculations. The first calculation normalizes the time axis of the transitional segment and the time axis of the preceding segment to a standard interval of zero to one relative position, corresponding them one-to-one according to their relative positions, and then assigns the transitional segment... Subtracting the actual vibration parameters of the preceding segment at the same relative position from the actual vibration parameters at each relative position, and then calculating the difference point by point, yields the second physical residual sequence between the transition segment and the preceding segment. The same method is then used to calculate the difference point by point between the transition segment and the following segment, resulting in the second physical residual sequence between them. Subsequently, the total energy of the two second physical residual sequences (the sum of the squares of the differences at all sampling points) is calculated. The magnitudes of the two total energies are compared, and the preset operating condition type of the adjacent actual time segment corresponding to the smaller total energy is taken as the preset operating condition type of the transition segment. When a transition segment has only one valid adjacent actual time segment in time, the preset operating condition type of that adjacent actual time segment is directly determined as the preset operating condition type of the transition segment, while the other end, lacking an adjacent segment, is not included in the comparison. The mechanism is as follows: the vibration characteristics of a transitional segment between two stable operating conditions are necessarily closer to the stable operating condition at one end. This is because the switching of operating conditions is a continuous and gradual process rather than an instantaneous jump. The vibration characteristics of the first half of the transitional segment are closer to the previous operating condition, and the second half are closer to the second operating condition. The smaller the overall vibration difference with the latter end indicates which end the overall vibration characteristics of the transitional segment are closer to. By normalizing the time axes of the two segments and subtracting them point by point according to their relative positions, the problem of time stamp misalignment caused by the difference in segment length is eliminated, allowing for quantitative comparison of vibration similarity between segments of different lengths. A smaller total residual energy means that the vibration waveforms of the two segments are more similar overall. Therefore, it is physically reasonable to assign the operating condition type of the corresponding adjacent segment to the transitional segment. This method does not require calling a physical model for forward simulation. It can complete the operating condition classification of the transitional segment using only actual vibration data. The computational cost is low and the implementation is simple. At the same time, it ensures that the transitional segment has a clear operating condition type to rely on when calling health baseline parameters and calculating physical residuals, avoiding the problem of being unable to participate in diagnosis due to missing operating conditions.

[0057] In S130, for each actual time segment, each time step within the segment is traversed first, and three types of features are extracted: For temperature features, a linear fit is performed on the actual temperature values ​​of the longer neighborhoods before and after the time step, and the slope of the fit is taken as the actual temperature change rate; for rotational speed features, the absolute value of the actual rotational speed at the time step is taken as the actual rotational speed absolute value, and the sign of the actual rotational speed value is taken as the actual rotational speed direction. A linear fit is performed on the actual rotational speed values ​​of the shorter neighborhoods before and after the time step, and the slope is taken as the actual rotational speed change rate; for load features, the absolute value of the actual load at the time step is directly taken as the actual load absolute value. Then, the three types of features for each time step are compared one by one with the first, second, and third ranges corresponding to each preset working condition in the preset working condition mapping table. If the temperature feature falls into the first range, the rotational speed feature falls into the second range, and the load feature falls into the third range, the working condition matching score is obtained. The preset working condition with the highest matching score is taken as the preset working condition corresponding to that time step. Time steps with matching scores all below the threshold are marked as working condition uncertainty points. After all time steps are determined, the number of valid time steps corresponding to each preset working condition within the segment is counted. When the number of time steps for a certain preset working condition is greater than or equal to the third threshold, the preset working condition is determined as the preset working condition of the actual time series segment. When the number of time steps for all preset working conditions is less than the third threshold, it indicates that the segment is in a working condition transition process. At this time, the preset working conditions of the preceding and following actual time series segments that are temporally adjacent to the segment are searched and recorded as the preceding and following working conditions, respectively. The segment is marked as a transition state segment and a dual baseline selection strategy is adopted in the subsequent physical residual calculation.

[0058] The preset operating condition mapping table is calibrated before system deployment by combining equipment rated parameters with statistical analysis of healthy operating data. First, based on the crane's design rated parameters (rated speed, rated load, transmission ratios of each mechanism) and typical operating procedures, the initial version of the mapping table is determined, including the range of absolute speed, direction of speed, range of speed change rate, and range of absolute load for each preset operating condition. Then, using historical operating data collected during the initial operation of the equipment or after a comprehensive overhaul to confirm its healthy state, statistical analysis is performed on the actual temperature change rate, actual absolute speed, actual speed change rate, and actual absolute load under each preset operating condition. The distribution range of quantities under healthy conditions (such as the average value plus or minus a certain number of standard deviations) serves as the boundary of the corresponding range, and the range of the initial version is corrected and refined. The first range of temperature characteristics is determined based on the statistical distribution of the rate of temperature change of each operating condition under healthy conditions. The second range of speed characteristics is determined based on the joint statistical distribution of the absolute value, direction, and rate of change of speed. The third range of load characteristics is determined based on the statistical distribution of the absolute value of load. After the mapping table is calibrated, it is stored in the system. It can be periodically reviewed and corrected as equipment operating data accumulates. If the actual operating characteristics of a certain operating condition continue to deviate from the original range, the range is updated to ensure that the mapping table always reflects the true operating characteristics of the equipment.

[0059] The mechanism of this scheme is as follows: each preset operating condition of the crane gearbox has a unique combination of characteristics in terms of temperature, speed, and load. The absolute value and direction of the speed determine the motion state of gear meshing, the speed change rate reflects the dynamic transition process, the absolute value of the load determines the tooth surface contact stress and vibration amplitude level, and the temperature change rate reflects the balance between heat generation and heat dissipation under that operating condition. Therefore, by matching these three types of characteristics with the preset range, the operating condition type of each time step can be accurately determined. The use of percentage statistics at the segment level instead of direct output step-by-step is because the instantaneous switching of operating conditions and short-term transients can cause individual... If the characteristics of a time step deviate from the dominant operating condition, transient interference can be filtered out and the dominant operating condition characteristics of the segment can be extracted by statistically counting the number of time steps of the same preset operating condition and comparing them with a third threshold. When the number of time steps of all operating conditions in a segment is less than the third threshold, it indicates that the segment itself is in a transition process between two stable operating conditions, and its operating characteristics do not meet the typical conditions of any single operating condition. At this time, its transition attribute is defined by the preset operating conditions of adjacent segments, and a dual baseline selection method is used in subsequent diagnosis. This avoids baseline mismatch caused by forced classification and ensures that the transition segment can still participate in fault diagnosis without being discarded.

[0060] Taking the hoisting mechanism gearbox of a bridge crane with a rated speed of 1,000 revolutions per minute and a rated load of 50 tons as an example, the uniform speed hoisting condition in the preset working condition mapping table is defined as follows: the first range of temperature change rate is 0.1 to 0.5 degrees Celsius per minute, the second range of speed characteristic is 800 to 1,050 revolutions per minute, the direction is forward rotation, the speed change rate is -10 to +10 revolutions per minute per second, and the third range of load characteristic is 5 to 50 tons; the hoisting acceleration condition is defined as the temperature change rate is 0.2 to 0.8 degrees Celsius per minute, the absolute value of speed is 200 to 800 revolutions per minute, the direction is forward rotation, the speed change rate is greater than +20 revolutions per minute per second, and the load is 5 to 50 tons; the stable operation condition is defined as the temperature change rate is -0.2 to 0.1 degrees Celsius, the absolute value of speed is 0 to 50 revolutions per minute, the direction is stationary, the speed change rate is close to zero, and the load is 0 to 2.5 tons. For a 10-second actual time segment, features are extracted step-by-step and matched with a mapping table. Assuming the segment has 20,000 time steps, 17,000 time steps match for uniform lift, 2,000 time steps match for lift acceleration, and 1,000 are uncertain points. If the third threshold is set to 15,000 time steps, the number of time steps for uniform lift exceeds the third threshold, and the preset condition for the segment is determined to be uniform lift. If another 2-second segment has 4,000 time steps, 2,000 match for uniform lift, 1,800 match for lift braking, and 200 are uncertain points, and the number of time steps for both conditions is below the third threshold, the segment is marked as a transition state. The uniform lift of the preceding segment and the lift braking of the following segment are taken as the preceding and following conditions, respectively. During subsequent diagnosis, the baseline parameters of the two conditions are used to generate predicted vibrations and calculate residuals. The one with the smaller residual energy is taken as the effective baseline.

[0061] Prior to S140, this embodiment also includes the following method for establishing a health baseline database, specifically:

[0062] The historical operating data of the crane gearbox under healthy conditions is acquired. The historical operating data includes historical speed time-series data, historical temperature time-series data, historical load time-series data, and historical vibration time-series data. Based on the historical speed, historical temperature, and historical load at each time step in the historical operating data, the historical vibration time-series data is divided into multiple continuous historical vibration segments, and the type of preset working condition corresponding to each historical vibration segment is determined. Based on the preset range of working condition parameters, the working condition parameters corresponding to each preset working condition are determined using a particle swarm optimization algorithm to form the health baseline library.

[0063] In this embodiment, based on the above scheme, the establishment process of the health baseline library is as follows: First, during the initial operation of the crane gearbox or the period after a comprehensive overhaul when the equipment is confirmed to be fault-free, historical operating data under healthy conditions is collected, including historical vibration time-series data collected synchronously at a high frequency of 20 kHz to 50 kHz, as well as synchronous historical speed, historical temperature, and historical load time-series data. The collection period must cover all typical operating conditions of the crane, and each operating condition must contain at least several complete healthy operating cycles. Then, the historical operating data is preprocessed in the same way as in the real-time diagnosis stage (vibration noise reduction, slow variable smoothing, and normalization). Based on the historical speed, historical temperature, and historical load at each time step, the operating condition switching time step is determined according to the same preset rules as in the real-time stage. The historical vibration time-series data is divided into multiple continuous historical vibration segments. Then, the preset operating condition type corresponding to each historical vibration segment is determined according to the same feature extraction and preset operating condition mapping table matching method as in the real-time stage. After removing transitional segments, segments with insufficient duration, and segments with unqualified data quality, the remaining healthy vibration segments are grouped according to the preset operating condition types, with each operating condition type corresponding to a group of healthy vibration segments.In the working condition parameter identification stage, for each preset working condition type, the historical rotational speed time series and historical load time series of all healthy vibration segments under that type are used as inputs, and the corresponding historical vibration time series is used as the fitting target. Based on the particle swarm optimization algorithm, the optimal working condition parameters are searched within the preset range of the working condition parameters. The specific process is as follows: First, the preset range of the working condition parameters to be identified (meshing stiffness reference and meshing damping reference) is set. This range is derived from the upper and lower bounds of design parameters such as the elastic modulus of gear material, tooth surface contact area, and typical damping ratio of the transmission system, ensuring that the search space is limited to a physically reasonable range; then, the initial... A particle swarm is constructed, where each particle represents a set of candidate operating condition parameters (i.e., a set of values ​​for meshing stiffness and meshing damping). The particles are randomly distributed within a preset range, and the number of particles is set according to the parameter dimensions (usually twenty to fifty). The initial velocity of each particle is also randomly set. Next, a fitness evaluation is performed on each particle. The meshing stiffness and damping corresponding to that particle are substituted into the gearbox lumped parameter dynamics model. Forward simulation is performed using the historical rotational speed and historical load of each healthy vibration segment under this operating condition as input, generating the predicted vibration time series for each segment. The predicted vibration is then subtracted point-by-point from the historical vibration of the same segment. The residual sequence is obtained, and the sum of the total energy of the three residuals of all segments is calculated as the fitness value of the particle. The smaller the fitness value, the better the vibration predicted by the model under that set of parameters matches the actual vibration of a healthy particle. Then, particle swarm iteration is performed. Each particle adjusts its velocity and position based on its own historical best position and the global best position of the population. The velocity update takes into account both individual experience (approaching its own historical best) and population experience (approaching the global best). If a particle exceeds the preset range of the working parameters after the position update, it is pulled back to the boundary. After each round of updates, the fitness values ​​of all particles are recalculated, and the fitness value of each particle is updated. The historical optimal position and the global optimal position of the population are determined. Iteration continues until the global optimal fitness value changes less than the preset convergence threshold for several consecutive rounds, or the maximum number of iterations is reached. At this time, the meshing stiffness and meshing damping values ​​corresponding to the global optimal position of the population are the optimal operating parameters for this preset operating condition type. The optimal operating parameters for each preset operating condition type are associated with the operating condition type number and stored together with the fixed parameters (transmission error, bearing support stiffness, moment of inertia, shaft torsional stiffness) determined by the design drawings and stored in the healthy baseline library to form a structured baseline parameter library with the preset operating condition type as the main key.The mechanism is as follows: the vibration response of the gearbox under healthy conditions is determined by its physical parameters (meshing stiffness, meshing damping, etc.) and external excitation (speed, load). The particle swarm optimization algorithm searches globally within a physically reasonable preset range to minimize the difference between the model's predicted vibration and the actual healthy vibration. Essentially, it identifies physical parameters reflecting the health status of the equipment from healthy vibration data. Since the search range is constrained by physical priors, the identification results will not have physically unreasonable values. Moreover, the parameters identified independently for each working condition can accurately reflect the differences in tooth surface contact state and damping characteristics caused by different loads and speeds under that working condition, thus providing an accurate parameter benchmark for generating working condition-specific healthy vibration predictions in the subsequent real-time diagnosis stage.

[0064] Based on this, S140 performs an index retrieval in the health baseline library using the preset operating condition type number determined for each actual time segment as the primary key, retrieves the two identification parameters, meshing stiffness reference and meshing damping reference, from the corresponding record, and reads them together with the fixed parameters (transmission error, bearing support stiffness, moment of inertia, shaft torsional stiffness) stored in the record into memory as the operating condition parameters for that actual time segment.

[0065] In S150, for each actual time-series segment, the operating parameters of that segment (meshing stiffness reference, meshing damping reference, and fixed parameters such as transmission error, bearing support stiffness, moment of inertia, and shaft torsional stiffness) obtained from the healthy baseline library are substituted into the gearbox model to complete the model instantiation. The instantiated model has a fixed structure and parameter values ​​corresponding to the operating conditions of that segment. Subsequently, the actual rotational speed time-series data and actual load time-series data of that actual time-series segment are used as external excitation inputs to the model, where the actual rotational speed serves as the kinematic input or driving constraint for each shaft, and the actual load serves as the external resistance torque acting on the output shaft. The model performs time-stepped forward simulation with the same time step as the vibration sampling interval, starting from the beginning of the segment. Each time step calculates internal forces such as gear meshing force and bearing support force based on the current rotational speed and load, and then solves for the acceleration, velocity, and displacement responses of each concentrated mass. The vibration acceleration response at a specified location (such as the input shaft bearing housing, output shaft bearing housing, and meshing zone housing wall) is output as the predicted vibration parameter for that time step. The process continues step by step until the end of the segment. The predicted vibration parameters of all time steps are arranged in chronological order to obtain the predicted vibration time series data for that specified location. The predicted vibration time series data is of the same length as the actual vibration time series data for that segment and the timestamps correspond one-to-one. When there are multiple specified locations, the model outputs the predicted vibration parameters for each location simultaneously at each time step, forming separate predicted vibration time series data.

[0066] The gearbox model adopts a lumped parameter dynamic model. The establishment process is as follows: First, based on the transmission structure of the crane gearbox, the system is discretized into several lumped elements, including the rotational inertia of each gear, the torsional stiffness of each transmission shaft, the time-varying meshing stiffness and meshing damping of the gear pairs, the transmission error excitation caused by gear manufacturing and installation errors, the bearing support stiffness and support damping of the shaft system, and the equivalent mass and equivalent stiffness of the gearbox at a specified measuring point. Each lumped element corresponds to one physical degree of freedom, and the elements are connected by elasticity and damping to form a multi-degree-of-freedom vibration system. Second, based on the mechanical connection relationship between the elements, the dynamic differential equation of the system is established. The equation includes a time-varying meshing stiffness term determined by the rotational speed (meshing stiffness changes periodically with the meshing position of the gear teeth), a displacement excitation term determined by the transmission error, and an external resistance torque term determined by the load. The model includes inertial terms, elastic restoring force terms, and damping dissipation terms for each degree of freedom. Fixed parameters in the model (moment of inertia, shaft torsional stiffness, transmission error, and bearing support stiffness) are directly determined by the gearbox design drawings, gear precision grades, and bearing model manuals, and do not change with operating conditions. Parameters in the model that change with operating conditions (meshing stiffness reference and meshing damping reference) are provided by a healthy baseline library, and different values ​​are taken under different operating conditions to reflect the influence of load and rotational speed on the tooth surface contact state and damping characteristics. The model uses numerical integration methods (such as the fourth-order Runge-Kutta method) to solve stepwise in the time domain. Given the rotational speed and load time sequence input, the vibration acceleration response of each measuring point can be output. The model solves quickly (the simulation of a segment is completed in milliseconds), and the parameters have clear physical meanings. They can be identified and calibrated through health data, making it suitable for real-time generation of healthy vibration predictions embedded in the fault diagnosis process.

[0067] In S160, for each actual time segment, the predicted vibration time series data generated by the gearbox model is aligned point-by-point with the synchronously acquired actual vibration time series data according to the timestamp. Since the predicted vibration time series data is generated on the same time axis as the actual vibration with the same sampling interval, the k-th sampling point of both naturally corresponds to the same moment, and no additional interpolation alignment is required. For each vibration channel at a specified location, the actual vibration parameter of the channel at each time step is subtracted from the predicted vibration parameter at the same time step and the same specified location. The subtraction operation is performed point-by-point to obtain the first physical residual sequence of the channel. When there are multiple specified locations, each channel is calculated independently to form multiple first physical residual sequences. During the point-by-point subtraction process, validity filtering is performed. If the actual vibration parameter or predicted vibration parameter of a certain sampling point exceeds the sensor range or the reasonable range of the model output, the point is marked as invalid, and the adjacent valid points are used as the invalid points. The validity value is replaced by linear interpolation. If the number of consecutive invalid points exceeds 10% of the total number of points in a segment, the segment is marked as having unqualified residual quality and will not be included in the subsequent deep learning model diagnosis. In the first physical residual sequence, the actual vibration and the predicted vibration under healthy conditions should be highly consistent. The residual should be close to the background noise level and have no significant periodic components. When a fault occurs, the actual vibration is superimposed with the abnormal vibration component caused by the fault excitation. This component is retained and relatively amplified in the residual, showing characteristics related to the fault type (such as tooth surface wear manifested as increased energy in the meshing frequency sideband, pitting manifested as periodic impact, broken teeth manifested as large-amplitude single impact, and bearing spalling manifested as high-frequency impact). Therefore, the first physical residual sequence is essentially the abnormal vibration component remaining after removing the expected vibration component determined by the rotational speed and load under healthy conditions, providing the subsequent deep learning model with physically meaningful fault enhancement feature input.

[0068] In S170, for each actual time-series segment, the first physical residual sequence of that segment, the actual vibration time-series data, and the segment's identifier are jointly input into the deep learning model. The model employs a three-branch parallel encoding architecture: the first branch takes the actual vibration time-series data as input and extracts the time-domain statistical features, frequency-domain features, and time-frequency joint features from the original vibration signal through a multi-layer one-dimensional convolutional or temporal convolutional network to capture the fault characterization of the vibration signal itself; the second branch takes the first physical residual sequence as input and extracts the abnormal component features from the physical residual through a convolutional encoding network with the same structure. Since the healthy expected vibration component has been removed from the residual, this branch focuses on the deviation features caused by the fault; the third branch takes the segment's identifier as input, which includes discrete information such as the segment's preset working condition type number, segment duration, and segment's position number within a specified time period. The feature vectors are converted into low-dimensional dense vectors through an embedding layer to encode the operating condition context information. The feature vectors output from the three branches are concatenated into a unified high-dimensional feature vector in the fusion layer, and then subjected to nonlinear transformation through one or more fully connected layers. Finally, the output layer outputs the probability distribution of each fault type corresponding to the segment. The probability distribution includes the probability values ​​of the fault-free state and the specified fault types such as tooth surface wear, tooth surface pitting, tooth breakage, bearing spalling, gear eccentricity, and shaft loosening. The sum of all probability values ​​is one. The category with the highest probability in the probability distribution is taken as the first initial fault diagnosis result of the actual time series segment. When the highest probability category is fault-free, the conclusion of fault-free is output. When the highest probability category is a specified fault type, the conclusion of the existence of that type of fault is output. At the same time, the highest probability value and the probability distribution of each category are recorded as a reference for diagnostic confidence for subsequent temperature and speed verification.

[0069] The deep learning model is trained using supervised learning before deployment. The training data consists of two parts: healthy state segments and fault state segments. The healthy state segments are derived from historical operational data collected during the establishment of the health baseline library. Each healthy segment is preprocessed, divided into operating conditions, and physical residuals are calculated to form a three-branch input sample, labeled as "no fault." The fault state segments are obtained through the following methods: first, real fault case data accumulated throughout the equipment's entire lifecycle (fault types confirmed by manual inspection); second, fault data collected on a gearbox test bench by artificially implanting faults (such as simulating pitting during EDM on tooth surfaces, simulating tooth breakage through tooth root grooving, and simulating spalling through scratches on the bearing outer ring); and third, fault data generated by simulation based on gearbox dynamics models with different degrees and types of fault excitation parameters injected. Each fault sample is labeled with a corresponding specified fault type. During training, all samples are divided into training and validation sets according to a preset ratio. The model is initialized with random parameters on a test set and a validation set. During forward propagation, each sample is encoded using a three-branch system, features are fused, and a fully connected layer is used to transform the model and output a probability distribution of the fault type. The difference between the model's output probability distribution and the true label is calculated using the cross-entropy loss function. The weights of each layer of the model are updated by backpropagation using a gradient descent optimization algorithm (such as the Adam optimizer). During training, samples are input in batches and parameters are updated iteratively batch by batch. After each round of training, the classification accuracy and loss value are calculated on the validation set. Training is stopped when the validation set accuracy no longer improves or the loss value converges for several consecutive rounds, and the model weights are saved. After training, the generalization performance of the model is evaluated on the test set, including the recognition accuracy, recall, and confusion matrix of various fault types. If the recognition rate of a certain type of fault is too low, training samples of that type of fault are added and the model is retrained. The final model weight file is deployed to the diagnostic system for real-time inference and output of the first initial fault diagnosis result.

[0070] In S170, this is specifically achieved in the following way:

[0071] Based on the actual temperature, the temperature rise rate of each time step in the actual time segment is calculated; based on the actual rotational speed of each time step in the actual time segment, the rotational speed fluctuation rate of the actual time segment is calculated; when the maximum temperature rise rate is greater than the rise rate threshold of the preset operating condition of the actual time segment, or the rotational speed fluctuation rate is greater than the fluctuation rate threshold of the preset operating condition of the actual time segment, the second initial diagnosis result indicates that the first initial diagnosis result is unreliable.

[0072] For each actual time series segment, firstly, the temperature rise rate at each time step is calculated based on the actual temperature data of that segment: taking each time step as the center, the actual temperature values ​​of several adjacent time steps are linearly fitted, and the slope of the fitting is taken as the temperature rise rate of that time step. The temperature rise rate represents the increase in temperature per unit time. After traversing all time steps within the segment, the maximum value is taken as the maximum temperature rise rate of that segment. Secondly, the speed fluctuation rate is calculated based on the actual speed data of that segment: taking the actual speed values ​​of all time steps within the segment, the ratio of their standard deviation to the average speed is calculated as the speed fluctuation rate of that segment. The speed fluctuation rate represents the relative fluctuation of speed within that segment. Subsequently, according to the preset operating condition type of the actual time series segment, the data is retrieved from the health baseline library. The system reads the temperature rise rate threshold and fluctuation rate threshold corresponding to the operating condition type. The temperature rise rate threshold is the normal upper limit determined by the statistical distribution of the temperature rise rate of all healthy segments under this operating condition type during the establishment of the health baseline library. The fluctuation rate threshold is the normal upper limit determined by the statistical distribution of the speed fluctuation rate of all healthy segments under this operating condition type. The system compares the maximum temperature rise rate with the temperature rise rate threshold and the speed fluctuation rate with the fluctuation rate threshold. When the maximum temperature rise rate is greater than the temperature rise rate threshold or the speed fluctuation rate is greater than the fluctuation rate threshold, the second initial diagnosis result indicates that the first initial fault diagnosis result is unreliable. When the maximum temperature rise rate is less than or equal to the temperature rise rate threshold and the speed fluctuation rate is less than or equal to the fluctuation rate threshold, the second initial diagnosis result indicates that the first initial fault diagnosis result is reliable. The definition of a reliable fault diagnosis is as follows: the temperature rise rate and speed fluctuation rate of the actual time segment are both within the healthy normal range of its preset operating condition type, indicating that the segment's operating state is stable, and its thermal and speed states are consistent with healthy operating conditions. There is no distortion of the vibration signal caused by transient operating conditions, overload, drastic speed fluctuations, or abnormal heating. Therefore, the first initial fault diagnosis result of the deep learning model based on the vibration signal and physical residual output has a reliable physical background, and its judgment conclusion can be accepted. The definition of an unreliable fault diagnosis is as follows: the temperature rise rate or speed fluctuation rate exceeds the healthy normal range, indicating that the segment is in an abnormal operating state of thermal instability or drastic speed fluctuations. At this time, the vibration signal may contain non-fault abnormal components caused by unstable operating conditions. The deep learning model may misidentify such operating condition disturbances as faults. Therefore, the physical background of the first initial fault diagnosis result is unreliable, its judgment conclusion is not accepted, and it is not included in the statistics of the final diagnosis result.

[0073] S180 specifically includes the following solutions:

[0074] When all the second initial fault diagnosis results that characterize the credibility of the first initial fault diagnosis result indicate that at least two of the time segments have the same fault type, and the at least two of the time segments correspond to at least two of the preset working conditions, then the final fault diagnosis result indicates that the crane gearbox has a fault of the fault type with the highest damage level.

[0075] Collect the second initial diagnostic results of all actual time-series segments within a specified time period, and filter out segments that represent the reliability of the first initial fault diagnosis results. These reliable segments are then grouped and statistically analyzed according to the fault type in their first initial fault diagnosis results, forming a fault statistics table. The table records the number of segments corresponding to each fault type, the preset operating condition type to which these segments belong, the fault probability of each segment, and the damage level of each segment. Subsequently, a joint judgment condition check is performed on each fault type: First, check whether the number of reliable segments corresponding to the fault type is greater than or equal to two; second, check whether the number of preset operating condition types corresponding to these segments is greater than or equal to two (i.e., the same fault type is detected under at least two different preset operating conditions). When both conditions are met... If all conditions are met simultaneously, the fault type is determined to be a real fault; otherwise, it is determined to be a suspected misjudgment or a single anomaly and is not included in the final fault conclusion. For fault types that pass the joint judgment conditions, the highest damage level is found among all its corresponding reliable segments, and the highest damage level is taken as the final damage level of the fault type. If multiple fault types pass the joint judgment conditions, each fault type and its highest damage level are output separately. The final fault diagnosis result output includes the confirmed fault type, the corresponding highest damage level, the number of segments that detected the fault and the distribution of operating conditions, and the diagnosis details of each segment. If all reliable segments are fault-free or no fault type meets the joint judgment conditions, the final diagnosis result indicates that the crane gearbox is fault-free. The mechanism is as follows: a real gearbox failure is a continuous damage to the equipment structure. It will not only show abnormalities in a certain operating condition segment while being completely normal in other operating conditions. The same type of failure can be reliably detected under at least two different preset operating conditions, eliminating the possibility of misjudgment caused by interference from specific operating conditions or abnormal data in a single segment. Therefore, the real failure can be confirmed. The highest damage level is taken as the final level because the highest level reflects the most serious damage state of the equipment, and it is safest to use this as the basis for operation and maintenance decisions.

[0076] In addition to the above-mentioned "cross-condition multi-segment consistent anomaly" judgment scheme, the following equivalent joint judgment schemes can also be used to achieve the same purpose of suppressing false alarms: Scheme 1 is the same condition recurrence scheme, that is, when the same fault type recurs in at least three credible time-series segments of the same preset operating condition type, it is judged as a real fault. This scheme eliminates single transient interference by detecting multiple recurrences under the same operating condition; Scheme 2 is the time-series continuous occurrence scheme, that is, when the same fault type appears in at least two time-series consecutive credible segments, it is judged as a real fault. This scheme uses the time persistence characteristics of the fault to eliminate isolated anomalies; Scheme 3 is the fault probability weighted accumulation scheme, that is, the fault probability values ​​of the same fault type in all credible segments are summed, and when the cumulative probability exceeds a preset threshold, it is judged as a real fault. This scheme quantifies the strength of fault evidence through probability accumulation; Scheme 4 is the damage level escalation scheme, that is, when the damage level of the same fault type in different segments shows a progressively increasing trend, it is judged as a real fault. This scheme uses damage level escalation to eliminate isolated anomalies. The monotonicity of the progress confirms the fault evolution; Scheme 5 is the cross-measurement point consistency scheme, that is, when the same fault type is reliably detected in vibration channels at at least two different specified locations, it is determined to be a real fault. This scheme eliminates false alarms caused by single sensor faults through cross-verification of multiple measurement points; Scheme 6 is the historical batch joint scheme, that is, when only a single segment of the current batch detects a certain fault type, it searches whether the same fault type has been reliably detected under the same operating condition in the historical batch. If it has occurred in the past and the interval is within the normal aging cycle of the equipment, then two or more detections are jointly determined to be a real fault. This scheme eliminates occasional anomalies through the accumulation of evidence across the time dimension; The above schemes can be used individually or in combination (such as confirming the fault by satisfying any one scheme, or confirming the fault by satisfying multiple schemes at the same time). When used in combination, the false alarm rate can be further reduced. The core mechanism is to distinguish between continuous structural damage and occasional operating condition interference through cross-verification of evidence from multiple dimensions and multiple segments, thereby improving the reliability of the final diagnostic conclusion.

[0077] 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 crane gearbox fault diagnosis method described above.

[0078] 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.

[0079] 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.

[0080] 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.

[0081] 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 crane gearbox fault diagnosis method as described in the above embodiments.

[0082] 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.

[0083] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the crane gearbox fault diagnosis method described above. 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.

[0084] 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.

[0085] 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.

[0086] 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 diagnosing faults in a crane gearbox, characterized in that, The method includes: The actual rotational speed, actual temperature, and actual load of the crane gearbox at each time step within a specified period are obtained, as well as the actual vibration parameters of the crane gearbox at at least one specified location, to obtain actual rotational speed time-series data, actual temperature time-series data, actual load time-series data, and actual vibration time-series data at the specified location. Based on the actual rotational speed, actual temperature, and actual load of each time step, the specified time period is divided into at least one continuous actual time sequence segment; Based on the actual rotational speed, actual temperature, and actual load of each time step in the actual time sequence segment, the type of preset operating condition for each actual time sequence segment is determined, and the number of preset operating conditions is several. In a pre-established health baseline library, the operating parameters of each actual time segment of the crane gearbox in a healthy state are obtained; The actual rotational speed time series data, actual load time series data, and operating condition parameters of each actual time series segment are input into the gearbox model. The gearbox model generates the predicted vibration parameters of the specified position at each time step through forward simulation, so as to obtain the predicted vibration time series data of each specified position. The first physical residual sequence between the predicted vibration time series parameters and the actual vibration time series parameters at each time step is obtained by using the difference between the predicted vibration time series data and the actual vibration time series data. The first physical residual sequence, the actual vibration time series data, and the identifier of the actual time series segment are input into the deep learning model. The deep learning model is used to obtain the first initial fault diagnosis result of the crane gearbox when it is in each actual time series segment. The first initial fault diagnosis result includes whether the crane gearbox is faultless when it is in the corresponding actual time series segment, or whether there is a fault of a specified fault type. Based on the actual temperature and actual rotation speed of each time step in the actual time sequence segment, the first initial fault diagnosis result is verified to obtain a second initial fault diagnosis result. The second initial fault diagnosis result indicates whether the first initial fault diagnosis result is credible or unreliable. The final diagnosis result of the crane gearbox is obtained based on all the second initial fault diagnosis results that characterize the credibility of the first initial fault diagnosis result.

2. The method according to claim 1, characterized in that, The step of dividing the specified time period into at least one continuous actual time sequence segment based on the actual rotational speed, actual temperature, and actual load at each time step includes: Based on the actual rotational speed, actual temperature and actual load of the time step, and according to preset rules, the time step for switching operating conditions is determined among all the time steps. The actual timing segment is determined based on the time step of the operating condition switch; The preset rule indicates that when the rate of change of the actual rotational speed corresponding to a first specified number of consecutive time steps exceeds a first threshold, the last time step of the specified number of time steps is the time step for the change of operating conditions. The preset rule indicates that when the absolute value of the rate of change of the actual load corresponding to a second specified number of consecutive time steps exceeds a second threshold, the last time step of the specified number of time steps is the time step of the working condition switching. The preset rule indicates that when the direction of the actual rotational speed changes abruptly, the time step of the direction change is the time step of the operating condition switch.

3. The method according to claim 1, characterized in that, The step of determining the type of preset operating condition for each actual time segment based on the actual rotational speed, actual temperature, and actual load at each time step in the actual time segment includes: Based on the actual temperature of each time step in the actual time series segment, the actual temperature characteristics of each time step in the actual time series segment are obtained, and the actual temperature characteristics include the actual temperature change rate. Based on the actual rotational speed of each time step in the actual time sequence segment, the actual rotational speed characteristics of each time step in the actual time sequence segment are obtained. The actual rotational speed characteristics include the actual rotational speed change rate, the actual rotational speed direction, and the actual rotational speed absolute value. Based on the actual load of each time step in the actual time series segment, the actual load characteristics of each time step in the actual time series segment are obtained, and the actual load characteristics include the absolute value of the actual load; Based on the actual temperature characteristics, actual speed characteristics, and actual load characteristics of each time step in the actual time sequence segment, and based on a preset operating condition mapping table, a preset operating condition corresponding to each time step in the actual time sequence segment is determined. The preset operating condition mapping table includes the mapping relationship between the preset operating condition and the first range of actual temperature characteristics, the second range of actual speed characteristics, and the third range of actual load characteristics. When the number of time steps of the same preset condition is greater than or equal to the third threshold, the corresponding preset condition is the preset condition of the actual time segment. When the number of time steps of the same preset condition is less than the third threshold, the preset condition of the actual time segment is determined according to the preset conditions of the adjacent actual time segments.

4. The method according to claim 3, characterized in that, The step of determining the preset operating condition corresponding to each time step of the actual time segment based on the actual temperature characteristics, actual rotational speed characteristics, and actual load characteristics of each time step in the actual time segment, and based on a preset operating condition mapping table, includes: Based on the mapping relationship recorded in the preset working condition mapping table, the matching degree between the time step and the type of each preset working condition is obtained through the matching score of the actual temperature feature with the first range, the matching score of the actual speed feature with the second range, and the matching score of the actual load feature with the third range, as well as the fixed weights of the pre-calibrated actual temperature feature, actual speed feature, and actual load feature. When the highest matching degree is greater than or equal to the fourth threshold and the highest matching degree is unique, the type of the highest preset working condition is taken as the preset working condition corresponding to each time step.

5. The method according to claim 3, characterized in that, The step of determining the preset operating conditions of the actual time segment based on the preset operating conditions of adjacent actual time segments includes: When there are adjacent actual time segments at both ends of the actual time segment in the time sequence, the second physical residual sequence of the actual time segment and the two adjacent actual time segments is calculated. The second physical residual sequence is obtained by the difference between the actual vibration parameter of the time step of the actual time segment and the actual vibration parameter of the time step of the adjacent actual time segment at the same position in the time sequence. The type of the preset operating condition of the adjacent actual time segment corresponding to the smaller value of the two second physical residual sequences is the type of the preset operating condition of the actual time segment. When only one end of the actual timing segment has the adjacent actual timing segment in the timing sequence, the type of the preset operating condition of the corresponding adjacent actual timing segment is the type of the preset operating condition of the actual timing segment.

6. The method according to claim 1, characterized in that, Before obtaining the operating parameters of each actual time segment of the crane gearbox in a healthy state from a pre-established health baseline library, the method further includes: Acquire historical operating data of the crane gearbox in a healthy state, including historical speed time series data, historical temperature time series data, historical load time series data, and historical vibration time series data; Based on the historical rotational speed, historical temperature, and historical load of each time step in the historical operating data, the historical vibration time series data is divided into multiple continuous historical vibration segments, and the type of preset working condition corresponding to each historical vibration segment is determined. Based on the preset range of operating parameters, the operating parameters corresponding to each preset operating condition are determined using the particle swarm optimization algorithm to form the health baseline library.

7. The method according to claim 1, characterized in that, The second initial diagnosis result is obtained by verifying the first initial fault diagnosis result based on the actual temperature and actual rotational speed at each time step in the actual time sequence segment, including: Based on the actual temperature, the temperature rise rate at each time step in the actual time segment is calculated; The rotational speed fluctuation rate of the actual time series segment is calculated based on the actual rotational speed at each time step in the actual time series segment. When the maximum temperature rise rate is greater than the rise rate threshold of the preset operating condition of the actual time segment, or the speed fluctuation rate is greater than the fluctuation rate threshold of the preset operating condition of the actual time segment, the second initial diagnosis result indicates that the first initial diagnosis result is unreliable.

8. The method according to claim 1, characterized in that, The final diagnosis result of the crane gearbox is obtained based on all second initial fault diagnosis results that are sequentially consecutive in time and characterize the reliability of the first initial fault diagnosis result, including: When all the second initial fault diagnosis results that characterize the credibility of the first initial fault diagnosis result indicate that at least two of the time segments have the same fault type, and the at least two of the time segments correspond to at least two of the preset working conditions, then the final fault diagnosis result indicates that the crane gearbox has a fault of the fault type with the highest damage level.

9. 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 8.

10. 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 8.

Images