Fault monitoring method and server for electromechanical servo system

Abstract

The invention discloses a fault monitoring method and a server applied to an electromechanical servo system, belonging to the technical field of electromechanical servo fault analysis. The method comprises: determining target light and heavy load state vector chains respectively corresponding to X embedded chip operation records to be analyzed corresponding to X pre-screened suspected fault events; extracting Y target cross jump state vector chains from the X target light and heavy load state vector chains according to trend change threshold characteristic values; respectively transmitting the Y target cross jump state vector chains to a fault diagnosis and discrimination network to generate Y initial cross abnormal monitoring labels; extracting Z target cross abnormal monitoring labels from the Y initial cross abnormal monitoring labels according to the cross trend change threshold characteristic values; determining the pre-screened suspected fault events corresponding to the Z target cross abnormal monitoring labels as Z target fault diagnosis events and reporting them to a host computer system, so that the fault monitoring reliability and safety of the entire electromechanical servo system can be improved.

Description

Technical Field

[0001] The invention belongs to the technical field of electromechanical servo fault analysis, and in particular relates to a fault monitoring method and a server applied to an electromechanical servo system. Background Art

[0002] In the field of electromechanical servo systems, traditional fault diagnosis methods often face many challenges. On the one hand, the operating states of the power drive unit, motor, reducer and other parts of the system are complex and changeable, and the characteristics they show under different load conditions are quite different, which makes it difficult for traditional single-state analysis methods to fully and accurately capture fault information. For example, some faults may not be obvious under light load, but will cause serious problems under heavy load, and the existing analysis methods may not be able to comprehensively consider the difference between light and heavy loads, resulting in missed or false alarms in fault detection. On the other hand, the existing fault diagnosis methods lack effective data processing and screening mechanisms when facing complex electromechanical servo systems. A large amount of operating data contains many interference factors, which makes it easy for the real fault signal to be submerged, making it difficult to accurately judge the occurrence and type of the fault. In addition, the preliminary judgment of the fault often lacks a further precise screening mechanism, which is easily affected by short-term interference and misjudgment, thereby affecting the reliability and safety of the entire system. Summary of the invention

[0003] The present invention provides a fault monitoring method and a server applied to an electromechanical servo system, which can solve or partially solve the technical problems involved in the above-mentioned background technology.

[0004] The present invention provides a fault monitoring method applied to an electromechanical servo system, which is applied to a fault monitoring server. The method comprises: respectively mining light and heavy load state vectors for X embedded chip operation records to be analyzed corresponding to X pre-screened suspected fault events, and obtaining a target light and heavy load state vector chain corresponding to each of the embedded chip operation records to be analyzed; extracting Y target cross jump state vector chains from the X target light and heavy load state vector chains according to a trend change threshold characteristic value, wherein Y is an integer not less than 1 and not greater than X; respectively transmitting the Y target cross jump state vector chains to a fault diagnosis and discrimination network, and generating Y initial cross abnormal monitoring labels through the fault diagnosis and discrimination network; extracting Z target cross abnormal monitoring labels from the Y initial cross abnormal monitoring labels according to the cross trend change threshold characteristic value, wherein Z is an integer not less than 1 and not greater than Y; determining the pre-screened suspected fault events corresponding to the Z target cross abnormal monitoring labels as Z target fault diagnosis events, and reporting the Z target fault diagnosis events to a host computer system.

[0005] The present invention provides a fault monitoring server, comprising at least one processor and a memory; the memory stores computer execution instructions; the at least one processor executes the computer execution instructions stored in the memory, so that the at least one processor executes the above method.

[0006] The present invention provides a readable storage medium, on which a program or instruction is stored, and the program or instruction implements the steps of the above method when executed by a processor.

[0007] In the present invention, the method includes determining the target light and heavy load state vector chains respectively corresponding to the X embedded chip operation records to be analyzed corresponding to the X preliminary screening suspected fault events; extracting Y target cross jump state vector chains from the X target light and heavy load state vector chains according to the trend change threshold characteristic value; respectively transmitting the Y target cross jump state vector chains to the fault diagnosis and discrimination network to generate Y initial cross abnormal monitoring labels; extracting Z target cross abnormal monitoring labels from the Y initial cross abnormal monitoring labels according to the cross trend change threshold characteristic value; determining the preliminary screening suspected fault events corresponding to the Z target cross abnormal monitoring labels as Z target fault diagnosis events and reporting them to the upper computer system, so that the reliability and safety of the entire electromechanical servo system can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 The present invention provides a flow chart of a fault monitoring method for an electromechanical servo system.

[0009] Figure 2 This is a structural schematic diagram of a fault monitoring server provided by the present invention. DETAILED DESCRIPTION

[0010] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0011] It should be noted that the formulas in the specific implementation manner have been converted into plain text form, and those skilled in the art can clearly and completely restore the corresponding formula contents based on this.

[0012] The terms "first", "second", etc. in the present invention are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present invention can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are generally of the same type, and the number of objects is not limited. For example, the first object can be one or more. In addition, in the present invention, "and / or" means at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0013] Figure 1 A fault monitoring method applied to an electromechanical servo system is shown, and applied to a fault monitoring server. The method includes the following steps 101 to 105.

[0014] Step 101 : mining light and heavy load state vectors of X embedded chip operation records to be analyzed corresponding to X preliminarily screened suspected fault events, respectively, to obtain a target light and heavy load state vector chain corresponding to each embedded chip operation record to be analyzed.

[0015] In the fault monitoring of electromechanical servo systems, it is necessary to analyze the operation records of embedded chips. Since the operating states of electromechanical servo systems are different under light and heavy load conditions, it is necessary to digitally discretize the continuous sensor monitoring signals. In this way, complex operating signals can be converted into analyzable state vector chains, providing basic data for subsequent fault diagnosis.

[0016] For example, there are (X=5) suspected fault events in the target electromechanical servo system, and each event corresponds to an embedded chip operation record to be analyzed. For each operation record, the continuous sensor monitoring signal, such as the motor speed signal, the power drive unit current signal, and the reducer torque signal, etc.

[0017] Taking the motor speed signal as an example, let its original continuous signal be (v(t)), and use equal time intervals (△t) for sampling to obtain a discretized speed sequence (v(n△t)), (n=0, 1, 2, cdots). The same sampling operation is performed on the current signal (i(t)) of the power drive unit and the torque signal (T(t)) of the reducer.

[0018] Then, these discretized signals are combined into state vectors according to adaptive rules. For example, in the case of light load, the state vector is defined (vec{S}_L=[v(n△t)_L, i(n△t)_L, T(n△t)_L]), and in the case of heavy load, the state vector is defined (vec{S}_H=[v(n△t)_H, i(n△t)_H, T(n△t)_H]). Over time, these state vectors form a light-heavy load state vector chain. For each embedded chip operation record to be analyzed, the corresponding target light-heavy load state vector chain can be obtained.

[0019] Step 102: extract Y target cross-jump state vector chains from the X target light-load and heavy-load state vector chains according to the trend change threshold characteristic value, wherein Y is an integer not less than 1 and not greater than X.

[0020] In this step, the trend change threshold characteristic value is a key parameter used to filter out the state vector chain with special state changes. It reflects the boundary between normal and abnormal system operation status. For example, if the rate of change of the motor speed exceeds a certain set trend change threshold characteristic value, it may mean that the system has an abnormal situation.

[0021] For example, the trend change threshold characteristic value is set to (α=100) rpm / s (revolutions per minute per second) for the speed change rate, (β=5) A / s (ampere per second) for the current change rate, and (γ=20) Nm / s (Newton meter per second) for the torque change rate.

[0022] Screen from (X=5) target light and heavy load state vector chains. For the adjacent state vectors (vec{S}_k) and (vec{S}_{k+1}) in each state vector chain, calculate the rate of change of speed, current and torque. If there is a jump in a state vector chain that satisfies at least one of the conditions (vertv((k+1)△t)-v(k△t)vert / △t>α), (verti((k+1)△t)-i(k△t)vert / △t>β), (vertT((k+1)△t)-T(k△t)vert / △t>γ), it is marked as a potential cross-jump state vector chain. After screening, (Y=3) target cross-jump state vector chains are obtained.

[0023] Step 103: respectively input the Y target cross-jump state vector chains into the fault diagnosis and discrimination network, and generate Y initial cross-abnormal monitoring labels through the fault diagnosis and discrimination network.

[0024] In the present invention, the fault diagnosis and discrimination network is a trained neural network model, which can receive the target cross-jump state vector chain as input and output the corresponding initial cross-abnormal monitoring label. The structure of the network can be a multi-layer perceptron (MLP), for example, having an input layer, a hidden layer and an output layer. The number of neurons in the input layer corresponds to the dimension of the target cross-jump state vector chain. For example, if the dimension of the target cross-jump state vector chain is (m), then the input layer has (m) neurons. There can be multiple hidden layers, for example, two hidden layers are set, the first hidden layer has (n_1) neurons, and the second hidden layer has (n_2) neurons. The output layer has one neuron, and the output value is 0 or 1, 0 means normal, and 1 means abnormal.

[0025] The (Y=3) target cross-jump state vector chains are respectively input into the fault diagnosis and discrimination network. Taking one of the target cross-jump state vector chains: (vec{S}_{jump}=[v_{jump}(n△t), i_{jump}(n△t), T_{jump}(n△t)]) as an example, the values ​​of its components are respectively input into the input layer neurons of the fault diagnosis and discrimination network.

[0026] The forward propagation calculation is performed through the weight matrix and activation function (such as ReLU function) in the network. For example, the weight matrices (W_1), (W_2) and (W_3) of the network correspond to the connection weights from the input layer to the first hidden layer, from the first hidden layer to the second hidden layer, and from the second hidden layer to the output layer. After calculation, a value is obtained in the output layer. If the value is greater than 0.5, the initial cross-abnormal monitoring label is 1, indicating abnormality; if it is less than or equal to 0.5, the initial cross-abnormal monitoring label is 0, indicating normality. This operation is performed for each target cross-jump state vector chain, and finally (Y=3) initial cross-abnormal monitoring labels are generated.

[0027] Step 104: extract Z target cross-abnormal monitoring labels from Y initial cross-abnormal monitoring labels according to the cross-trend change threshold characteristic value, wherein Z is an integer not less than 1 and not greater than Y.

[0028] In this step, the cross trend change threshold feature value is a further screening criterion based on the initial cross anomaly monitoring label. It can more accurately determine which anomalies are target faults that really need attention. For example, in some cases, although initially judged to be abnormal, it may be caused by short-term interference. These false anomalies can be excluded by the cross trend change threshold feature value.

[0029] For example, the duration of the cross trend change threshold feature value for the initial cross abnormal monitoring tag is set to (t_{min}=5)s (seconds). For the abnormal conditions corresponding to (Y=3) initial cross abnormal monitoring tags, check their duration. If the duration of the abnormal condition corresponding to an initial cross abnormal monitoring tag is less than (t_{min}), it is regarded as a short-term interference and is not considered as a target fault.

[0030] After screening, for example, (Z=2) target cross-anomaly monitoring labels are obtained, which means that among the (Y=3) initial anomalies, one is caused by short-term interference and is excluded.

[0031] Step 105: initially screen suspected fault events corresponding to the Z target cross-abnormal monitoring tags as Z target fault diagnosis events, and report the Z target fault diagnosis events to the host computer system.

[0032] In detail, the initial screening suspected fault events corresponding to (Z=2) target cross-abnormal monitoring tags are determined as target fault diagnosis events. The faulty units corresponding to these target fault diagnosis events include at least one of the power drive unit, the motor unit, and the reducer unit. For example, one target fault diagnosis event may be caused by a short circuit in the winding of the motor unit, and another may be caused by damage to the power tube in the power drive unit.

[0033] Furthermore, the target fault diagnosis event (Z=2) is reported to the host computer system. After receiving these fault events, the host computer system can perform subsequent fault shutdown or protection actions according to the pre-set strategy. For example, for a motor unit failure, the host computer can issue a command to stop the motor to prevent further damage; for a power drive unit failure, the host computer can cut off the power supply or adjust the drive signal and other operations.

[0034] In the technical solution recorded in the above steps 101 to 105, the X embedded chip operation records to be analyzed belong to the same target electromechanical servo system; the light and heavy load state vector mining is used to perform digital discretization processing on the continuous sensor monitoring signals corresponding to the X embedded chip operation records to be analyzed; the fault unit corresponding to the target fault diagnosis event includes at least one of a power drive unit, a motor unit and a reducer unit.

[0035] In order to facilitate a clearer and more complete implementation of the above technical solution, the target light and heavy load state vector chain, the target cross jump state vector chain, the initial cross abnormality monitoring label, the target cross abnormality monitoring label, and the target fault diagnosis event are further introduced below.

[0036] The target light and heavy load state vector chain is obtained by digitally discretizing the continuous sensor monitoring signal corresponding to the operation record of the embedded chip in the electromechanical servo system. It is a quantitative representation of the system under different light and heavy load working states.

[0037] For example, for an electromechanical servo system, its sensor monitoring signals include the speed of the motor, the current of the power drive unit, and the torque of the reducer. Under light load, a state vector may be (vec{S}_L=[v(n△t)_L, i(n△t)_L, T(n△t)_L]), where (v(n△t)_L) is the speed sampling value of the motor at time (n△t) under light load, (i(n△t)_L) is the current sampling value of the power drive unit at time (n△t) under light load, and (T(n△t)_L) is the torque sampling value of the reducer at time (n△t) under light load. As time (n) changes, these state vectors form a light load state vector chain. The heavy load state vector chain is constructed in a similar way, except that it corresponds to the sampling value of the system under heavy load.

[0038] The target light and heavy load state vector chain provides basic data for subsequent fault diagnosis. Since the normal operating state of the system is different under light and heavy load conditions, potential fault signs can be found by analyzing the change trends of these state vector chains. For example, under normal circumstances, the range of motor speed changes, current size, and torque fluctuations under light load have certain rules. If the data in the state vector chain deviates from these rules, it may indicate that the system is faulty.

[0039] The target cross-jump state vector chain is extracted from the target light-heavy load state vector chain according to the trend change threshold characteristic value. It reflects the large-scale jump situation in the system operation state.

[0040] For example, the trend change threshold characteristic value is set to (α=100) rpm / s (revolutions per minute per second) for the motor speed change rate, (β=5) A / s (ampere per second) for the current change rate, and (γ=20) Nm / s (Newton meter per second) for the torque change rate. When the adjacent state vectors (vec{S}_k) and (vec{S}_{k+1}) in a state vector chain meet at least one of the following conditions: (vertv((k+1)△t)-v(k△t)vert / △t>α), (verti((k+1)△t)-i(k△t)vert / △t>β), (vertT((k+1)△t)-T(k△t)vert / △t>γ), the state vector chain may be marked as a potential cross-jump state vector chain, and the target cross-jump state vector chain is obtained after further screening.

[0041] This jump may be caused by a fault in the system, such as a sudden load change of the motor (possibly a jam of mechanical components), a sudden damage to the components of the power drive unit (causing a sudden change in current), or a fault inside the reducer (causing an abnormal jump in torque). By identifying the target cross-jump state vector chain, the scope of troubleshooting can be narrowed to those time periods and states where obvious abnormal changes occur.

[0042] The initial cross-abnormal monitoring label is generated by passing the target cross-jump state vector chain into the fault diagnosis discrimination network. The fault diagnosis discrimination network can be a neural network with a multi-layer perceptron (MLP) structure.

[0043] For example, the target cross-jump state vector chain (vec{S}_{jump}=[v_{jump}(n△t), i_{jump}(n△t), T_{jump}(n△t)]) is input into the fault diagnosis discrimination network. The number of neurons in the input layer of the network corresponds to the dimension of the state vector chain, for example, (m) neurons. There can be multiple hidden layers, such as setting two hidden layers, the first hidden layer has (n_1) neurons, the second hidden layer has (n_2) neurons, and the output layer has one neuron. Forward propagation calculations are performed through the weight matrices (W_1), (W_2) and (W_3) in the network and the activation function (such as the ReLU function). The output layer obtains a value. If the value is greater than 0.5, the initial cross-abnormal monitoring label is 1, indicating abnormality; if it is less than or equal to 0.5, the initial cross-abnormal monitoring label is 0, indicating normality.

[0044] The initial cross-abnormal monitoring tag is a preliminary judgment on whether the target cross-jump state vector chain is abnormal. It provides a preliminary basis for further accurate screening of target faults, but may contain some misjudgments caused by factors such as short-term interference.

[0045] The target cross-abnormal monitoring label is extracted from the initial cross-abnormal monitoring label based on the cross-trend change threshold feature value. For example, the duration of the cross-trend change threshold feature value for the initial cross-abnormal monitoring label is set to (t_{min}=5)s (seconds). For the abnormal situation corresponding to the initial cross-abnormal monitoring label, check its duration. If the duration of the abnormal situation corresponding to a certain initial cross-abnormal monitoring label is less than (t_{min}), it is regarded as a short-term interference and is not considered as a target fault. After screening, the target cross-abnormal monitoring label is obtained.

[0046] The target cross-abnormal monitoring label more accurately reflects the real fault situation in the system. It eliminates the misjudgment caused by short-term interference, making the fault diagnosis result more reliable.

[0047] The target fault diagnosis event is determined by the initial screening of suspected fault events corresponding to the target cross-abnormal monitoring tags. The faulty units corresponding to these events include at least one of the power drive unit, the motor unit, and the reducer unit. For example, through the previous analysis, if a target cross-abnormal monitoring tag corresponds to a continuous abnormal decrease in motor speed, and is associated with abnormal current fluctuations in the power drive unit, it is determined through analysis that the motor winding is short-circuited or the power tube in the power drive unit is partially damaged, which constitutes a target fault diagnosis event. The target fault diagnosis event clarifies the specific circumstances of the fault in the system and provides an accurate basis for the host computer system to carry out targeted processing.

[0048] Based on the above content, the specific processing methods of the host computer system for different target fault diagnosis events are as follows.

[0049] 1. Power drive unit failure

[0050] 1. Power tube damage

[0051] If the target fault diagnosis event is determined to be a power tube failure in the power drive unit, the host computer system will first cut off the control signal to the power drive unit to prevent the fault from further expanding. For example, the host computer sends a command to the control circuit of the power drive unit to stop the output of the PWM (pulse width modulation) signal.

[0052] Then, the host computer system will record the time when the fault occurred, related monitoring data (such as current and voltage when the fault occurred), and send an alarm signal to the operator. The alarm signal can be displayed on the monitoring interface, and can also notify the relevant maintenance personnel through voice or text messages.

[0053] The host computer system can also try to perform some preliminary repair operations according to pre-set strategies, such as restarting part of the circuit of the power drive unit (if the circuit has a self-repair function), or adjusting other related parameters to maintain some functions of the system (if the system design allows).

[0054] 2. Driving circuit short circuit

[0055] When a short circuit fault is detected in the drive circuit, the host computer system will immediately cut off the power supply to the entire power drive unit to prevent damage to other components caused by excessive short-circuit current, such as affecting the motor or power supply system.

[0056] The host computer system will record the circuit parameters related to the fault in detail, such as the voltage drop and current change near the short-circuit point, so that maintenance personnel can quickly locate the fault point. At the same time, the host computer system will notify the maintenance personnel to carry out repairs, and according to the redundancy design of the system, determine whether it is possible to switch to the backup power drive unit (if any) to maintain the basic operation of the system.

[0057] (II) Motor unit failure

[0058] 1. Motor winding short circuit

[0059] For the target fault diagnosis event of motor winding short circuit, the host computer system will stop sending drive signals to the motor to stop the motor from running. This is because continued operation may cause the motor to overheat, further damage the motor or even cause a safety accident.

[0060] The host computer system will collect the motor's operating data before the fault occurs, such as speed, torque, temperature, etc., as well as the changes in related parameters when the fault occurs. These data will help maintenance personnel determine the severity and possible causes of the winding short circuit.

[0061] The host computer system can also control the heat dissipation system (if any) related to the motor to continue working for a period of time according to the overall layout of the system to prevent the motor from causing greater damage due to residual heat after it stops running. At the same time, it notifies maintenance personnel to repair the motor or replace the winding.

[0062] 2. Motor bearing wear

[0063] When the motor bearing wear is diagnosed, the host computer system will reduce the load or speed of the motor (if the system allows) to relieve the pressure on the bearing. For example, if the motor is used to drive a mechanical device, the host computer can adjust the working mode of the device to reduce the torque demand on the motor.

[0064] The host computer system will continuously monitor the operating status of the motor, such as vibration amplitude, noise level, etc., and record relevant data. It will also issue a warning to the operator, indicating that the bearing may need to be replaced so that maintenance personnel can arrange maintenance work in time.

[0065] (III) Reducer unit failure

[0066] 1. Reducer gear wear

[0067] If the target fault diagnosis event is reducer gear wear, the host computer system will adjust the motor output torque and speed to adapt to the new state of the reducer. For example, the motor output speed is reduced, while the torque is appropriately increased to keep the output power of the entire electromechanical servo system basically stable.

[0068] The host computer system will closely monitor the temperature changes of the reducer, because gear wear may cause increased friction, which in turn causes the temperature to rise. If the temperature exceeds the safety threshold, the host computer system will stop the system and notify maintenance personnel to repair the reducer or replace the gears.

[0069] At the same time, the host computer system will record the operating data related to gear wear, such as the input-output torque ratio and speed changes at different stages. These data are very important for analyzing the progress and causes of gear wear.

[0070] 2. The reducer is stuck inside

[0071] When the reducer is determined to be stuck inside, the host computer system will immediately stop the motor to prevent the motor from being overloaded and damaged. If the motor continues to run, the reducer cannot transmit normally and the motor load will increase sharply.

[0072] The host computer system will record the relevant parameters when the fault occurs, such as the sudden increase in the motor current and the sudden decrease in speed. This information will help maintenance personnel determine the cause of the jam. The host computer system will also notify maintenance personnel to inspect and repair the reducer, and it may be necessary to disassemble the reducer to remove the jammed parts or repair the damaged parts.

[0073] In the actual application process, the fault diagnosis and discrimination network is the key invention point in steps 101 to 105, and its performance quality is crucial. Based on this, the present invention further designs an optimization method for the fault diagnosis and discrimination network, which exemplarily includes the following steps 201 to 206.

[0074] Step 201: perform light and heavy load state vector mining on each of the X embedded chip operation record samples to obtain a light and heavy load state vector chain sample corresponding to each of the embedded chip operation record samples, wherein each of the light and heavy load state vector chain samples corresponds to a priori training annotation, and X is an integer not less than 1.

[0075] Step 202: extract Y cross-jump state vector chain samples from the X light-load and heavy-load state vector chain samples according to the trend change threshold characteristic value, wherein Y is an integer not less than 1 and not greater than X.

[0076] Step 203: respectively input the Y cross-jump state vector chain samples into the fault diagnosis and discrimination network, and generate Y cross-abnormal monitoring labels through the fault diagnosis and discrimination network.

[0077] Step 204: respectively input the Y cross-jump state vector chain samples into the associated fault decision algorithm, and generate Y associated fault decision viewpoints through the associated fault decision algorithm.

[0078] Step 205 : According to the abnormal monitoring comparison restriction feature, the fault analysis error is determined for the Y cross abnormal monitoring tags and the Y associated fault decision viewpoints to obtain a first fault analysis comparison error.

[0079] Step 206: Optimize the neural network weights of the fault diagnosis and discrimination network according to the first fault analysis comparison error.

[0080] For step 201, when optimizing the fault diagnosis and discrimination network, the data foundation for training and optimization must first be obtained. By mining the light and heavy load state vectors of the embedded chip operation record samples, the complex operation records can be converted into state vector chain samples that can be used for subsequent analysis. This mining process discretizes the continuous operation information to better capture the characteristics of the system under different load conditions.

[0081] For example, there are (X=10) embedded chip operation record samples. For each sample, the sensor monitoring signals, such as the motor speed, the current of the power drive unit, and the torque of the reducer, are processed. Taking the motor speed signal as an example, the original signal is (v(t)), and it is sampled at a fixed time interval (△t=0.1)s to obtain a discretized speed sequence (v(n△t)), (n=0, 1, 2, cdots). Similarly, the current signal (i(t)) and the torque signal (T(t)) are sampled. Then, these discretized signals are combined according to the adaptive rules to form a state vector. Under light load conditions, the state vector is (vec{S}_L=[v(n△t)_L, i(n△t)_L, T(n△t)_L]), and under heavy load conditions, (vec{S}_H=[v(n△t)_H, i(n△t)_H, T(n△t)_H]). Over time, these state vectors form a light-heavy load state vector chain sample corresponding to each embedded chip operation record sample. In addition, each light-heavy load state vector chain sample corresponds to a prior training annotation, which may contain known system status information, such as whether it is running normally, whether there is a certain type of fault, etc., for subsequent training and evaluation.

[0082] For step 202, the trend change threshold characteristic value is the key to screening the cross-jump state vector chain samples. Its setting is based on the understanding of the range of changes of various parameters under the normal operation of the system. For example, for the motor speed change rate, the trend change threshold characteristic value is set to (α=80) rpm / s (revolutions per minute per second), for the current change rate, it is set to (β=4) A / s (ampere / second), and for the torque change rate, it is set to (γ=15) Nm / s (Newton meter / second).

[0083] Screening is performed from (X=10) light and heavy load state vector chain samples. For each adjacent state vector (vec{S}_k) and (vec{S}_{k+1}) in the state vector chain sample, the rate of change of the speed, current and torque is calculated. If a state vector chain sample has a jump condition that satisfies at least one of the following conditions: (vertv((k+1)△t)-v(k△t)vert / △t>α), (verti((k+1)△t)-i(k△t)vert / △t>β), (vertT((k+1)△t)-T(k△t)vert / △t>γ), it is marked as a potential cross-jump state vector chain sample. After screening, (Y=6) cross-jump state vector chain samples are obtained. These cross-jump state vector chain samples represent the abnormal state changes that may exist in the system operation, and are important data for subsequent analysis and optimization of the fault diagnosis and discrimination network.

[0084] For step 203, the fault diagnosis and discrimination network is a designed neural network for analyzing the cross-jump state vector chain sample and generating a cross-abnormal monitoring label. For example, the fault diagnosis and discrimination network adopts a multi-layer perceptron (MLP) structure, and the number of input layer neurons is determined according to the dimension of the cross-jump state vector chain sample. For example, if the state vector dimension of the cross-jump state vector chain sample is (m=3) (such as motor speed, current, torque), then the input layer has (m=3) neurons.

[0085] There are two hidden layers, the first hidden layer has (n_1=5) neurons, and the second hidden layer has (n_2=3) neurons. The output layer has one neuron, and the output value is 0 or 1, 0 means normal, and 1 means abnormal.

[0086] The (Y=6) cross-jump state vector chain samples are respectively passed into the fault diagnosis and discrimination network. Taking one of the cross-jump state vector chain samples (vec{S}_{jump}=[v_{jump}(n△t), i_{jump}(n△t), T_{jump}(n△t)]) as an example, the values ​​of its components are respectively input into the input layer neurons of the fault diagnosis and discrimination network. The weight matrices (W_1), (W_2) and (W_3) in the network correspond to the connection weights from the input layer to the first hidden layer, from the first hidden layer to the second hidden layer, and from the second hidden layer to the output layer, respectively, and the activation function (such as ReLU function) are used for forward propagation calculation. For example, the element values ​​of the weight matrix (W_1) are randomly generated during initialization, and the value range is between ([-0.5, 0.5]), for example (W_1=begin{bmatrix}0.1&-0.3&0.2-0.4&0.1&0.30.2&-0.1&0.4end{bmatrix}). After calculation, a value is obtained at the output layer. If the value is greater than 0.5, the cross-abnormal monitoring label is 1, indicating abnormality; if it is less than or equal to 0.5, the cross-abnormal monitoring label is 0, indicating normality. This operation is performed for each cross-jump state vector chain sample, and finally (Y=6) cross-abnormal monitoring labels are generated.

[0087] For step 204, the associated fault decision algorithm is an algorithm for analyzing the cross-jump state vector chain sample and generating an associated fault decision view. Its input is the cross-jump state vector chain sample, and its output is the associated fault decision view, which may be a preliminary judgment on the fault type, fault location or fault severity.

[0088] The (Y=6) cross-jump state vector chain samples are respectively passed to the associated fault decision algorithm. For example, the associated fault decision algorithm may make judgments based on the rule base. If the motor speed in the cross-jump state vector chain sample suddenly drops and the current rises sharply, according to the rules in the rule base, the associated fault decision algorithm may generate an associated fault decision view that the motor may have a winding short circuit fault or a sudden load increase fault. For each cross-jump state vector chain sample, the associated fault decision algorithm will generate a corresponding associated fault decision view based on its internal rules and logic.

[0089] For step 205, the abnormal monitoring comparison restriction feature is a standard for measuring the difference between the cross-abnormal monitoring label and the associated fault decision viewpoint. It can be a threshold or a set of rules for judging the consistency between the two.

[0090] For (Y=6) cross-abnormal monitoring labels and (Y=6) associated fault decision viewpoints, the fault analysis error is determined according to the abnormal monitoring comparison restriction feature. For example, the abnormal monitoring comparison restriction feature stipulates that if the cross-abnormal monitoring label is 0 (indicating normal) and the associated fault decision viewpoint is that there is a fault, then it is considered that there is an error; or if the cross-abnormal monitoring label is 1 (indicating abnormality) but the associated fault decision viewpoint is that there is no fault, it is also considered that there is an error. By comparing each pair of cross-abnormal monitoring labels and associated fault decision viewpoints, the proportion of the number of errors to the total number is counted to obtain the first fault analysis comparison error. For example, after comparison, it is found that there are (2) pairs with errors, then the first fault analysis comparison error is (2 / 6≈0.33).

[0091] For step 206, the neural network weights of the fault diagnosis and discrimination network are optimized according to the obtained first fault analysis comparison error. The purpose of the optimization is to reduce the error and improve the accuracy of the fault diagnosis and discrimination network. The back propagation algorithm is usually used to adjust the weights of the neural network.

[0092] Taking the multi-layer perceptron structure in step 203 as an example, the value of the loss function is calculated according to the first fault analysis comparison error. For example, the mean square error (MSE) is used as the loss function, (MSE=frac{1}{Y}sum_{i=1}^{Y}(y_i-hat{y}_i)^2), where (y_i) is the true value corresponding to the associated fault decision point of view (which can be determined based on the prior training annotation), and (hat{y}_i) is the cross-abnormal monitoring label value output by the fault diagnosis discrimination network. The weight matrices (W_1), (W_2) and (W_3) are derived according to the loss function to obtain the gradient value. Then, the weight matrix is ​​updated according to a certain learning rate (for example, (eta=0.01)), (W_{new}=W_{old}-etafrac{partialMSE}{partialW}). Through multiple iterations of this process, the weight matrix is ​​continuously adjusted until the first fault analysis comparison error meets certain requirements, such as less than (0.1).

[0093] With such a design, the accuracy of the fault diagnosis and discrimination network can be improved through the optimization method of the fault diagnosis and discrimination network. First, representative data is obtained for training and evaluation through the mining of light and heavy load state vectors and the extraction of cross-jump state vector chain samples. Then, the associated fault decision viewpoint is generated with the help of the associated fault decision algorithm, and compared with the cross-abnormal monitoring label generated by the fault diagnosis and discrimination network to determine the fault analysis error. Finally, the weight of the fault diagnosis and discrimination network is optimized based on the error. This series of steps can effectively reduce misjudgment and improve the accuracy of fault diagnosis of components such as power drive units, motors and reducers in electromechanical servo systems, thereby improving the reliability and safety of the entire electromechanical servo system.

[0094] In some optional embodiments R1, after determining the fault analysis error of the Y cross-abnormal monitoring labels and the Y associated fault decision viewpoints according to the abnormal monitoring comparison restriction characteristics to obtain the first fault analysis comparison error, the method further includes: determining the fault analysis error of the Y cross-abnormal monitoring labels according to the abnormal monitoring causal restriction characteristics to obtain the fault tracing error; optimizing the neural network weights of the fault diagnosis and discrimination network according to the first fault analysis comparison error, including: optimizing the neural network weights of the fault diagnosis and discrimination network according to the first fault analysis comparison error and the fault tracing error.

[0095] In some other optional embodiments R2, the extracting Y cross-jump state vector chain samples from the X light and heavy load state vector chain samples according to the trend change threshold characteristic value includes: determining a scenario fault diagnosis network according to the prior training annotations, and passing the light and heavy load state vector chain samples corresponding to the prior training annotations into the scenario fault diagnosis network, and generating scenario fault characteristic values ​​corresponding to the light and heavy load state vector chain samples through the scenario fault diagnosis network; determining, from the X light and heavy load state vector chain samples, the light and heavy load state vector chain samples whose scenario fault characteristic values ​​are not less than the trend change threshold characteristic values, to obtain Y cross-jump state vector chain samples. Furthermore, after the Y cross-jump state vector chain samples are respectively transmitted to the fault diagnosis and discrimination network and Y cross-abnormal monitoring labels are generated by the fault diagnosis and discrimination network, the method also includes: determining the fault analysis error of the Y cross-abnormal monitoring labels and the scenario fault feature values ​​according to the abnormal monitoring comparison restriction characteristics to obtain a second fault analysis comparison error; and optimizing the neural network weights of the fault diagnosis and discrimination network according to the second fault analysis comparison error.

[0096] In an alternative embodiment R3, the Y cross-jump state vector chain samples are respectively transmitted to the fault diagnosis and discrimination network, and after Y cross-abnormal monitoring labels are generated by the fault diagnosis and discrimination network, the method also includes: obtaining a priori cross-abnormal discrimination opinion corresponding to each of the embedded chip operation record samples; determining the fault analysis error of the Y cross-abnormal monitoring labels and the priori cross-abnormal discrimination opinion according to the abnormal monitoring comparison restriction characteristics to obtain a third fault analysis comparison error; and optimizing the neural network weights of the fault diagnosis and discrimination network according to the third fault analysis comparison error.

[0097] In the present invention, the following is an exemplary introduction to the above-mentioned Embodiment R1 to Embodiment R3.

[0098] 1. Example R1 Exemplary Description

[0099] 1. Determination of fault tracing error

[0100] 1. The significance and role of the causal restriction feature of abnormal monitoring

[0101] The anomaly monitoring causal restriction feature is an important basis for measuring the accuracy of cross-anomaly monitoring labels from the perspective of causality. It reflects the inherent connection logic between the fault phenomenon and the fault cause in the system, which plays a key role in determining the fault tracing error.

[0102] 2. Fault tracing error determination process

[0103] After obtaining the first fault analysis comparison error, the fault analysis error of (Y) cross-abnormal monitoring tags is determined according to the abnormal monitoring causal restriction feature to obtain the fault tracing error. For example, (Y=8), for each cross-abnormal monitoring tag, if it is marked as abnormal (for example, the value is 1) from the perspective of causality, but according to the known causal logic of the system (defined by the abnormal monitoring causal restriction feature), in the system state reflected by the current light-heavy load state vector chain sample, this abnormality should not occur (that is, according to the causal logic, it should be normal, for example, the value is 0), then it is considered that there is an error. For example, after analyzing (Y=8) cross-abnormal monitoring tags, it is found that (3) have such causal contradictions, then the fault tracing error is (3 / 8=0.375).

[0104] (II) Neural Network Weight Optimization Based on Two Errors

[0105] 1. Consideration of comprehensive error and optimization principle

[0106] The neural network weights of the fault diagnosis and discrimination network are optimized based on the first fault analysis comparison error and the fault tracing error. This is because the first fault analysis comparison error reflects the overall consistency difference between the fault diagnosis and discrimination network and the associated fault decision algorithm, while the fault tracing error reveals the problems of the fault diagnosis and discrimination network alone from the perspective of causality. The combination of the two can more comprehensively reflect the accuracy of the fault diagnosis and discrimination network, thereby optimizing the weights more accurately.

[0107] 2. Optimization Example

[0108] For example, the fault diagnosis and discrimination network is a multi-layer perceptron (MLP) structure, with an input layer of (m=4) neurons, corresponding to different parameters in the cross-jump state vector chain sample (such as motor speed, current, torque and temperature), two hidden layers, the first hidden layer has (n_1=6) neurons, the second hidden layer has (n_2=4) neurons, and the output layer is a neuron, outputting the cross-abnormal monitoring label (0 or 1). Let the weight matrix of the neural network be (W_1) (input layer to the first hidden layer), (W_2) (first hidden layer to the second hidden layer) and (W_3) (second hidden layer to the output layer).

[0109] When calculating the optimization, the first fault analysis comparison error (E_1) and the fault tracing error (E_2) are first considered comprehensively. For example, a comprehensive error (E=w_1E_1+w_2E_2) can be defined by weighted summation, where (w_1=0.6), (w_2=0.4) (the weights are determined based on the empirical evaluation of the impact of the two on the overall accuracy). Then the loss function is calculated based on the comprehensive error (E), for example, the cross entropy loss function (L=-[yln(hat{y})+(1-y)ln(1-hat{y})] is used, where (y) is the true label (0 or 1) determined based on the associated fault decision viewpoint and the abnormal monitoring causal restriction feature, and (hat{y}) is the cross abnormal monitoring label output by the fault diagnosis discriminant network.

[0110] Next, the weight matrices (W_1), (W_2), and (W_3) are differentiated according to the loss function to obtain the gradient value. For example, for (W_1), (frac{partialL}{partialW_1}) is calculated by the chain rule. Finally, the weight matrix is ​​updated according to a certain learning rate (eta=0.005), (W_{new}=W_{old}-etafrac{partialL}{partialW}). Through multiple iterations of this process, the weight matrix is ​​continuously adjusted until the comprehensive error (E) meets certain requirements, such as less than (0.1).

[0111] II. Example R2 Exemplary Description

[0112] 1. Obtaining the cross-jump state vector chain example

[0113] 1. Determination and function of scenario fault diagnosis network

[0114] The scenario fault diagnosis network is determined based on the prior training annotations. The prior training annotations contain rich system operation information, such as known failure modes in different operation scenarios, parameter ranges for normal operation, etc. The scenario fault diagnosis network is built based on this prior knowledge. It can analyze the light and heavy load state vector chain samples and extract the fault feature values ​​related to the scenario.

[0115] 2. The process of obtaining a cross-jump state vector chain sample

[0116] For example, (X=10), for each light and heavy load state vector chain sample, it is passed to the scenario fault diagnosis network. For example, the number of neurons in the input layer of the scenario fault diagnosis network is determined according to the dimension of the light and heavy load state vector chain sample, set to (p=5), the hidden layer has (q_1=4) neurons and (q_2=3) neurons, and the output layer outputs the scenario fault feature value.

[0117] Assume that the trend change threshold characteristic value is (T=0.5) (this value is a normalized value set according to the system performance requirements and experience, and is used to uniformly measure the scenario fault characteristic values ​​of different parameters). From (X=10) light and heavy load state vector chain samples, determine the light and heavy load state vector chain samples whose scenario fault characteristic values ​​are not less than the trend change threshold characteristic value, thereby obtaining (Y) cross-jump state vector chain samples. For example, after calculation by the scenario fault diagnosis network, the scenario fault characteristic value of a light and heavy load state vector chain sample is (0.6), which is greater than (T=0.5), then the light and heavy load state vector chain sample is determined as a cross-jump state vector chain sample.

[0118] (II) Determination and weight optimization of the second fault analysis comparison error

[0119] 1. Determination of the second fault analysis comparison error

[0120] After the (Y) cross-jump state vector chain samples are respectively passed into the fault diagnosis discriminant network to generate (Y) cross-abnormal monitoring labels, the fault analysis error of the (Y) cross-abnormal monitoring labels and the scene fault feature values ​​is determined according to the abnormal monitoring comparison restriction feature to obtain the second fault analysis comparison error. For example, the abnormal monitoring comparison restriction feature stipulates that if the cross-abnormal monitoring label is 0 (indicating normal), but the scene fault feature value is greater than the trend change threshold feature value (indicating that a fault may exist), it is considered that there is an error; or if the cross-abnormal monitoring label is 1 (indicating abnormality), but the scene fault feature value is less than the trend change threshold feature value (indicating that it may be normal), it is also considered that there is an error. For example, (Y=7), after comparison, it is found that there are (2) such inconsistencies, so the second fault analysis comparison error is (2 / 7≈0.286).

[0121] 2. Weight optimization based on the comparison error of the second fault analysis

[0122] The neural network weights of the fault diagnosis and discrimination network are optimized according to the second fault analysis comparison error. For example, the fault diagnosis and discrimination network adopts a multi-layer perceptron structure similar to the previous one, with (m=3) neurons in the input layer, (n_1=5) neurons and (n_2=3) neurons in the hidden layer, and one neuron in the output layer. The mean square error (MSE) is used as the loss function, (MSE=frac{1}{Y}sum_{i=1}^{Y}(y_i-hat{y}_i)^2), where (y_i) is the true label determined according to the scene fault feature value (0 or 1, determined by whether it is greater than the trend change threshold feature value), and (hat{y}_i) is the cross-abnormal monitoring label value output by the fault diagnosis and discrimination network.

[0123] According to the loss function, the weight matrices (W_1), (W_2) and (W_3) are differentiated to obtain the gradient value. For example, for (W_1), (frac{partialMSE}{partialW_1}) is calculated by the corresponding derivative formula. Then the weight matrix is ​​updated according to the learning rate (eta=0.01), (W_{new}=W_{old}-etafrac{partialMSE}{partialW}). Through multiple iterations of this process, the weight matrix is ​​continuously adjusted until the second fault analysis comparison error meets certain requirements, such as less than (0.1).

[0124] III. Example R3 Exemplary Description

[0125] 1. Determination of the third fault analysis comparison error

[0126] 1. Acquisition and function of prior cross-anomaly discrimination viewpoints

[0127] The prior cross-abnormal discrimination viewpoint is a pre-existing discrimination result about cross-abnormality for each embedded chip operation record sample. It may be obtained based on past experience, historical data or other relevant knowledge. After (Y) cross-jump state vector chain samples are passed into the fault diagnosis discrimination network to generate (Y) cross-abnormal monitoring labels, the prior cross-abnormal discrimination viewpoint is obtained to further analyze the accuracy of the fault diagnosis discrimination network.

[0128] 2. Process of determining the third fault analysis comparison error

[0129] For example (Y=9), for each cross-abnormal monitoring label and prior cross-abnormal discrimination viewpoint, the fault analysis error is determined according to the abnormal monitoring comparison restriction feature to obtain the third fault analysis comparison error. For example, the abnormal monitoring comparison restriction feature stipulates that if the cross-abnormal monitoring label is inconsistent with the prior cross-abnormal discrimination viewpoint (such as one is 0 and the other is 1), it is considered that there is an error. After comparison, it is found that there are (3) such inconsistent situations, so the third fault analysis comparison error is (3 / 9=1 / 3≈0.333).

[0130] 2. Weight optimization based on the third fault analysis comparison error

[0131] 1. Weight optimization process

[0132] The neural network weights of the fault diagnosis and discrimination network are optimized according to the third fault analysis comparison error. For example, the fault diagnosis and discrimination network is a multi-layer perceptron structure with an input layer (m=4) neurons (corresponding to different parameters in the cross-jump state vector chain sample), two hidden layers, the first hidden layer has (n_1=7) neurons, the second hidden layer has (n_2=5) neurons, and the output layer is a neuron.

[0133] The logarithmic loss function (L=-sum_{i=1}^{Y}[y_iln(hat{y}_i)+(1-y_i)ln(1-hat{y}_i)]) is used, where (y_i) is the prior cross-abnormal discrimination viewpoint (0 or 1), and (hat{y}_i) is the cross-abnormal monitoring label output by the fault diagnosis discrimination network. The weight matrices (W_1), (W_2), and (W_3) are derived according to the loss function to obtain the gradient value. For example, for (W_1), (frac{partialL}{partialW_1}) is calculated by the corresponding derivative formula. Then the weight matrix is ​​updated according to the learning rate (eta=0.008), (W_{new}=W_{old}-etafrac{partialL}{partialW}). Through multiple iterations of this process, the weight matrix is ​​continuously adjusted until the third fault analysis comparison error meets certain requirements, such as less than (0.1).

[0134] It can be seen that in Example R1, the neural network weights of the fault diagnosis and discrimination network are optimized by simultaneously considering the first fault analysis comparison error and the fault tracing error, and the accuracy of the fault diagnosis and discrimination network is evaluated from multiple angles. Not only the overall consistency with the associated fault decision algorithm is considered, but also an in-depth analysis is performed from the perspective of cause and effect, thereby more comprehensively improving the accuracy of the network.

[0135] In Example R2, a scenario fault diagnosis network is used to generate scenario fault feature values ​​to obtain cross-jump state vector chain samples, and the second fault analysis comparison error is determined based on its relationship with the cross-abnormal monitoring label for weight optimization, so that the fault diagnosis discrimination network can better adapt to the fault diagnosis needs in different scenarios and improve the ability to identify specific scenario faults.

[0136] In Example R3, the prior cross-abnormal discrimination viewpoint is introduced to determine the third fault analysis comparison error and optimize the weight, which makes full use of the existing prior knowledge, further improves the accuracy and reliability of the fault diagnosis and discrimination network, and helps to more accurately diagnose faults in the electromechanical servo system and ensure the normal operation of the system.

[0137] In a possible example, the scenario fault diagnosis network includes a power drive unit fault diagnosis subnet and a motor unit fault diagnosis subnet, and the scenario fault characteristic value includes a power drive fault characteristic value and a motor fault characteristic value; the scenario fault diagnosis network is determined according to the prior training annotation, and the light and heavy load state vector chain sample corresponding to the prior training annotation is transmitted to the scenario fault diagnosis network, and the scenario fault characteristic value corresponding to the light and heavy load state vector chain sample is generated through the scenario fault diagnosis network, including: when the prior training annotation is a power drive fault annotation, the power drive is determined Unit fault diagnosis subnet, and the light and heavy load state vector chain samples corresponding to the power drive fault annotation are passed into the power drive unit fault diagnosis subnet, and the power drive fault feature values ​​corresponding to the light and heavy load state vector chain samples are generated by the power drive unit fault diagnosis subnet; when the prior training annotation is a motor fault annotation, the motor unit fault diagnosis subnet is determined, and the light and heavy load state vector chain samples corresponding to the motor fault annotation are passed into the motor unit fault diagnosis subnet, and the motor fault feature values ​​corresponding to the light and heavy load state vector chain samples are generated by the motor unit fault diagnosis subnet. Then, determining, from the X light and heavy load state vector chain samples, a light and heavy load state vector chain sample in which the scenario fault characteristic value is not less than the trend change threshold characteristic value to obtain Y cross jump state vector chain samples, includes: determining, from the X light and heavy load state vector chain samples, a light and heavy load state vector chain sample in which the power drive fault characteristic value and the motor fault characteristic value are not less than the trend change threshold characteristic value to obtain Y cross jump state vector chain samples.

[0138] In the next step, when the prior training annotation is a motor fault annotation, the motor unit fault diagnosis subnet is determined, and the light and heavy load state vector chain samples corresponding to the motor fault annotation are passed into the motor unit fault diagnosis subnet. After the motor fault characteristic values ​​corresponding to the light and heavy load state vector chain samples are generated by the motor unit fault diagnosis subnet, the method also includes: determining the power drive fault analysis error according to the power drive fault annotation and the power drive fault characteristic value, and optimizing the neural network weights of the power drive unit fault diagnosis subnet according to the power drive fault analysis error; determining the motor fault analysis error according to the motor fault annotation and the motor fault characteristic value, and optimizing the neural network weights of the motor unit fault diagnosis subnet according to the motor fault analysis error.

[0139] In the above example, the scenario fault diagnosis network includes a power drive unit fault diagnosis subnet and a motor unit fault diagnosis subnet. This structural design is based on the importance and particularity of the power drive unit and motor unit in the fault diagnosis of the electromechanical servo system. Each subnet is specifically used to diagnose the fault of the corresponding unit, which improves the pertinence and accuracy of the fault diagnosis.

[0140] Furthermore, the generation mechanism of the scenario fault feature value is as follows: the scenario fault feature value includes the power drive fault feature value and the motor fault feature value. These feature values ​​are generated by passing the light and heavy load state vector chain samples into the corresponding subnets. Specifically, when there are prior training annotations, the corresponding light and heavy load state vector chain samples are passed into different subnets according to the type of annotation.

[0141] For example, there are a total of (X=10) prior training annotations, of which (X_1=4) are power drive fault annotations and (X_2=6) are motor fault annotations. For the light and heavy load state vector chain samples corresponding to the power drive fault annotations, they are passed to the power drive unit fault diagnosis subnet. This subnet may have a specific neural network structure, such as the input layer determines the number of neurons according to the dimension of the light and heavy load state vector chain samples, which is set to (m_1=5) neurons, and there are two hidden layers, the first hidden layer has (n_11=4) neurons, the second hidden layer has (n_12=3) neurons, and the output layer outputs the power drive fault feature value.

[0142] Similarly, for the light and heavy load state vector chain samples corresponding to the motor fault annotation, they are passed into the motor unit fault diagnosis subnet. The number of neurons in the input layer of the motor unit fault diagnosis subnet is set to (m_2=4) neurons, with two hidden layers, the first hidden layer has (n_21=3) neurons, the second hidden layer has (n_22=2) neurons, and the output layer outputs the motor fault feature value.

[0143] In addition, the acquisition of cross-jump state vector chain samples includes the following contents: In order to obtain cross-jump state vector chain samples, it is necessary to determine the light and heavy load state vector chain samples whose power drive fault characteristic values ​​and motor fault characteristic values ​​are not less than the trend change threshold characteristic values ​​from (X) light and heavy load state vector chain samples.

[0144] Assume that the trend change threshold characteristic value is (T=0.6) (this value is a normalized value determined according to the system performance requirements and a large amount of experimental data, and is used to uniformly measure whether the power drive fault characteristic value and the motor fault characteristic value have reached the degree of possible failure). For the power drive fault characteristic value output by the power drive unit fault diagnosis subnet and the motor fault characteristic value output by the motor unit fault diagnosis subnet, check the characteristic values ​​corresponding to the (X=10) light and heavy load state vector chain samples one by one.

[0145] For example, if the power drive fault characteristic value corresponding to a light or heavy load state vector chain sample is (0.7) and the motor fault characteristic value is (0.65), both of which are greater than (T=0.6), then the light or heavy load state vector chain sample is determined to be a cross-jump state vector chain sample. In this way, (Y) cross-jump state vector chain samples are obtained.

[0146] Next, an exemplary introduction to weight optimization of the power drive unit fault diagnosis subnetwork is given as follows.

[0147] 1. Determination of power drive fault analysis error

[0148] The power drive fault analysis error is determined based on the power drive fault annotation and the power drive fault characteristic value. For example, the power drive fault annotation specifies whether a certain light or heavy load state vector chain sample has a power drive unit fault under specific conditions (represented by 0 or 1, 0 for no fault and 1 for fault). The power drive fault characteristic value output by the power drive unit fault diagnosis subnet is a continuous value, such as (0.8).

[0149] An error calculation method, such as an absolute error calculation method, can be used. Let (y_{pd}) be the true value in the power drive fault annotation, and (hat{y}_{pd}) be the power drive fault characteristic value output by the power drive unit fault diagnosis subnet. Then the power drive fault analysis error (E_{pd}=verty_{pd}-hat{y}_{pd}vert). For example, (y_{pd}=1), then (E_{pd}=vert1-0.8vert=0.2).

[0150] 2. Error-based weight optimization

[0151] The neural network weights of the power drive unit fault diagnosis subnetwork are optimized according to the power drive fault analysis error. For example, the neural network weight matrix of the power drive unit fault diagnosis subnetwork is (W_{pd1}) (input layer to the first hidden layer), (W_{pd2}) (first hidden layer to the second hidden layer) and (W_{pd3}) (second hidden layer to the output layer).

[0152] The mean square error (MSE) is used as the loss function, (MSE_{pd}=frac{1}{N}sum_{i=1}^{N}(y_{pdi}-hat{y}_{pdi})^2), where (N) is the number of samples involved in the calculation (here is the number of samples corresponding to the power drive fault annotation), (y_{pdi}) is the true value in the power drive fault annotation, and (hat{y}_{pdi}) is the power drive fault feature value output by the power drive unit fault diagnosis subnet.

[0153] According to the loss function, the weight matrices (W_{pd1}), (W_{pd2}) and (W_{pd3}) are differentiated to obtain the gradient value. For example, for (W_{pd1}), (frac{partialMSE_{pd}}{partialW_{pd1}}) is calculated by the chain rule. Then the weight matrix is ​​updated according to the learning rate (eta=0.01), (W_{pd1, new}=W_{pd1, old}-etafrac{partialMSE_{pd}}{partialW_{pd1}}). Through multiple iterations of this process, the weight matrix is ​​continuously adjusted until the power drive fault analysis error meets certain requirements, such as less than (0.1).

[0154] Finally, an exemplary presentation of the weight optimization of the motor unit fault diagnosis subnetwork is as follows.

[0155] 1. Determination of motor fault analysis error

[0156] Similar to the power drive unit fault diagnosis subnet, the motor fault analysis error is determined based on the motor fault annotation and the motor fault characteristic value. Assume that the actual fault condition corresponding to a light or heavy load state vector chain sample in the motor fault annotation is (y_{me}) (0 or 1), and the motor fault characteristic value output by the motor unit fault diagnosis subnet is (hat{y}_{me}). Using the relative error calculation method, the motor fault analysis error (E_{me}=frac{verty_{me}-hat{y}_{me}vert}{y_{me}}) (when (y_{me}neq0)). For example, (y_{me}=1), (hat{y}_{me}=0.9), then (E_{me}=frac{vert1-0.9vert}{1}=0.1).

[0157] 2. Error-based weight optimization

[0158] The neural network weights of the motor unit fault diagnosis subnetwork are optimized according to the motor fault analysis error. The neural network weight matrix of the motor unit fault diagnosis subnetwork is set to (W_{me1}) (input layer to the first hidden layer), (W_{me2}) (first hidden layer to the second hidden layer) and (W_{me3}) (second hidden layer to the output layer).

[0159] The cross entropy loss function (L_{me}=-[y_{mei}ln(hat{y}_{mei})+(1-y_{mei})ln(1-hat{y}_{mei})]) is used, where (y_{mei}) is the true value in the motor fault annotation, and (hat{y}_{mei}) is the motor fault feature value output by the motor unit fault diagnosis subnetwork. The weight matrices (W_{me1}), (W_{me2}) and (W_{me3}) are derived according to the loss function to obtain the gradient value. For example, for (W_{me1}), (frac{partialL_{me}}{partialW_{me1}}) is calculated by the chain rule. Then the weight matrix is ​​updated according to the learning rate (eta=0.005), (W_{me1,new}=W_{me1,old}-etafrac{partialL_{me}}{partialW_{me1}}). Through multiple iterations of this process, the weight matrix is ​​continuously adjusted until the motor fault analysis error meets certain requirements, such as less than (0.05).

[0160] In this way, by dividing the scenario fault diagnosis network into a power drive unit fault diagnosis subnet and a motor unit fault diagnosis subnet, and generating corresponding fault characteristic values ​​respectively, the faults of different units can be diagnosed more accurately. When obtaining the cross-jump state vector chain sample, the relationship between the power drive fault characteristic value and the motor fault characteristic value and the trend change threshold characteristic value is comprehensively considered to improve the accuracy of screening potential fault states. Moreover, the fault analysis error is determined for the power drive unit fault diagnosis subnet and the motor unit fault diagnosis subnet respectively, and the weight is optimized, which further improves the diagnostic accuracy of the two subnets, helps to more accurately identify the faults of the power drive unit and the motor unit in the electromechanical servo system, and ensure the stable operation of the system.

[0161] Under some exemplary design ideas, the light and heavy load state vectors of each of the X embedded chip operation record samples are mined to obtain a light and heavy load state vector chain sample corresponding to each of the embedded chip operation record samples, including: obtaining a platform interaction description vector of an electromechanical servo system sample corresponding to the embedded chip operation record sample, wherein the platform interaction description vector is used to characterize the servo system state elements between each of the electromechanical servo system samples and the suspected fault event sample; according to the electromechanical servo system sample and the platform interaction description vector, the light and heavy load state vectors of the embedded chip operation record sample are mined to obtain a light and heavy load state vector chain sample corresponding to each of the embedded chip operation record samples.

[0162] It can be understood that the detailed description of this exemplary design concept may include the following two sections.

[0163] 1. Acquisition and Significance of Platform Interaction Description Vector

[0164] 1. Acquisition of platform interaction description vector

[0165] When mining the light and heavy load state vectors of embedded chip operation record samples, the platform interaction description vector of the electromechanical servo system sample must be obtained first. This vector is an important data structure used to characterize the servo system state elements between each electromechanical servo system sample and the suspected fault event sample. Its acquisition may involve integrating and analyzing a variety of information during the operation of the electromechanical servo system. For example, various operating parameters of the electromechanical servo system are collected through sensors, such as the speed, torque, voltage and current of the power drive unit of the motor, and related data generated when the system interacts with external devices or the environment, such as changes in load, input and output of control signals, etc.

[0166] For example, the operating data of the electromechanical servo system sample is stored in a data matrix (D), whose rows represent different time points and columns represent different parameters (such as speed, torque, voltage, current, etc.). This data matrix can be processed by a specific algorithm to extract information related to the suspected fault event sample, and then construct a platform interaction description vector. For example, the principal component analysis (PCA) algorithm is used to reduce the dimension of the data matrix (D) to obtain several main components, which can be used as part of the platform interaction description vector. Suppose the principal component vector obtained after PCA processing is (vec{v}=[v_1, v_2, v_3]), which may be the initial form of the platform interaction description vector.

[0167] 2. The significance of platform interaction description vector

[0168] The platform interaction description vector plays a key connecting role in the entire fault diagnosis process. It links the electromechanical servo system sample with the suspected fault event sample, providing an important basis for the subsequent light and heavy load state vector mining. Through this vector, we can have a more comprehensive understanding of the relationship between the system and the suspected fault event in different states, so as to more accurately mine the light and heavy load state vector chain samples. For example, if an element in the platform interaction description vector represents the correlation between a sudden change in load and a suspected fault event, then in the light and heavy load state vector mining process, we can focus on the impact of this factor on the system state vector.

[0169] 2. Mining of light and heavy load state vectors based on platform interaction description vectors

[0170] 1. Mining Principle

[0171] According to the electromechanical servo system sample and the platform interaction description vector, the light and heavy load state vector of the embedded chip operation record sample is mined. Specifically, based on the data in the embedded chip operation record sample and combined with the information in the platform interaction description vector, the state vector of the system under light and heavy load conditions is determined. For example, the embedded chip operation record sample contains the real-time speed data of the motor (r(t)), the real-time current data of the power drive unit (i(t)), and the real-time torque data of the reducer (T(t)).

[0172] An element in the platform interaction description vector may represent the weight of the impact of load changes on the system state under specific conditions. Let this weight be (w). If (w) is large, it means that the load change has a significant impact on the system state. When mining the light and heavy load state vector, the motor speed, current and torque data can be adjusted according to the value of (w). For example, when (w=0.8), for the speed data (r(t)), it can be converted into a state value (r'(t)) related to light and heavy loads according to the load changes. For example, the conversion formula is , where (△L) is the load change. Similarly, the current data (i(t)) and torque data (T(t)) are similarly transformed to obtain (i'(t)) and (T'(t)).

[0173] 2. Constructing a light and heavy load state vector chain example

[0174] After the above conversion, the processed motor speed (r'(t)), current (i'(t)) and torque (T'(t)) data are combined into a state vector in time sequence. In the case of light load, the state vector can be expressed as (vec{S}_L=[r'_L(n△t), i'_L(n△t), T'_L(n△t)]), and in the case of heavy load, the state vector is (vec{S}_H=[r'_H(n△t), i'_H(n△t), T'_H(n△t)]), where (n) represents the time point index and (△t) is the time interval. As time goes by, these state vectors constitute the light and heavy load state vector chain samples corresponding to each embedded chip operation record sample. For example, in a time series (n=0, 1, 2, cdots, 10), each time point has a corresponding light load or heavy load state vector, and these state vectors are connected in sequence to form a light and heavy load state vector chain sample.

[0175] It can be seen that by obtaining the platform interaction description vector, the servo system state elements between the electromechanical servo system sample and the suspected fault event sample are integrated into the light and heavy load state vector mining process. This method makes the mining of light and heavy load state vectors more accurate, and no longer relies solely on the original data in the embedded chip operation record sample. Using the information in the platform interaction description vector, such as the weight of the impact of load changes on the system state, the original data can be reasonably adjusted to construct a light and heavy load state vector chain sample that can better reflect the actual operating state of the system. This helps to more accurately screen out cross-jump state vector chain samples in the subsequent fault diagnosis process, improve the accuracy of the fault diagnosis discrimination network, and thus improve the reliability and effectiveness of the entire electromechanical servo system fault diagnosis.

[0176] In some preferred embodiments, when the abnormal monitoring comparison restriction feature is the quantized comparison information of the label vector; the fault analysis error is determined for the Y cross-abnormal monitoring labels and the Y associated fault decision viewpoints according to the abnormal monitoring comparison restriction feature to obtain a first fault analysis comparison error, including: according to the prior training annotations, from the Y cross-abnormal monitoring labels, the power drive cross-abnormal monitoring label corresponding to the power drive prior training annotation, the number of power drive abnormal events corresponding to the power drive cross-abnormal monitoring label, the motor unit cross-abnormal monitoring label corresponding to the motor unit prior training annotation, and the number of motor unit abnormal events corresponding to the motor unit cross-abnormal monitoring label; according to the prior training annotations, from the Y associated fault decision viewpoints, the power drive associated fault decision viewpoint corresponding to the power drive prior training annotation, the power drive associated fault decision viewpoint corresponding to the power drive The method comprises the following steps: determining the number of power drive associated abnormal events corresponding to the motor unit prior training annotation, the motor unit associated fault decision viewpoint corresponding to the motor unit prior training annotation, and the number of motor unit associated abnormal events corresponding to the motor unit associated fault decision viewpoint; determining the power drive cross feature according to the power drive cross abnormal monitoring label and the number of power drive abnormal events, and determining the motor unit cross feature according to the motor unit cross abnormal monitoring label and the number of motor unit abnormal events; determining the power drive associated feature according to the power drive associated fault decision viewpoint and the number of power drive associated abnormal events, and determining the motor unit associated feature according to the motor unit associated fault decision viewpoint and the number of motor unit associated abnormal events; performing fault analysis error determination on the power drive cross feature, the motor unit cross feature, the power drive associated feature and the motor unit associated feature to obtain the first fault analysis comparison error.

[0177] It can be understood that the first fault analysis comparison error determination based on label vector quantization comparison information includes three sub-steps: classification determination based on prior training annotations, feature determination, and fault analysis error determination.

[0178] 1. Classification Determination Based on Prior Training Annotations

[0179] 1. Determination of power drive related labels

[0180] According to the prior training annotations, the power drive cross-abnormal monitoring labels corresponding to the power drive prior training annotations and the corresponding number of power drive abnormal events are determined from the (Y) cross-abnormal monitoring labels. For example, the prior training annotations contain the operating status information of different parts of the system. For the power drive part, when (Y=10), after screening, it may be found that (Y_1=3) cross-abnormal monitoring labels are related to the power drive prior training annotations. For example, these (3) power drive cross-abnormal monitoring labels are (L_{pd1}), (L_{pd2}), and (L_{pd3}). The corresponding number of power drive abnormal events is obtained by counting the number of times marked as abnormal (for example, marked as 1 for abnormality) in a certain time series. Assume that the number of abnormal events corresponding to (L_{pd1}) is (n_{pd1}=5), the number of abnormal events corresponding to (L_{pd2}) is (n_{pd2}=3), and the number of abnormal events corresponding to (L_{pd3}) is (n_{pd3}=4).

[0181] Similarly, the power drive associated fault decision viewpoints corresponding to the power drive prior training annotations and the corresponding number of power drive associated abnormal events are determined from the (Y) associated fault decision viewpoints. For example, the power drive associated fault decision viewpoints obtained after screening are (D_{pd1}), (D_{pd2}), (D_{pd3}), and the corresponding numbers of power drive associated abnormal events are (m_{pd1}=4), (m_{pd2}=3), and (m_{pd3}=5).

[0182] 2. Determination of motor unit related labels

[0183] For motor units, the same method is used. Determine the motor unit cross-abnormal monitoring labels and the number of motor unit abnormal events corresponding to the motor unit prior training annotations from the (Y) cross-abnormal monitoring labels. For example, in (Y=10), there are (Y_2=4) motor unit cross-abnormal monitoring labels, namely (L_{me1}), (L_{me2}), (L_{me3}), (L_{me4}), and the corresponding number of abnormal events are (n_{me1}=2), (n_{me2}=3), (n_{me3}=4), and (n_{me4}=3).

[0184] Determine the motor unit associated fault decision viewpoints and the number of motor unit associated abnormal events corresponding to the motor unit prior training annotations from the (Y) associated fault decision viewpoints. For example, the motor unit associated fault decision viewpoints are (D_{me1}), (D_{me2}), (D_{me3}), (D_{me4}), and the corresponding numbers of associated abnormal events are (m_{me1}=3), (m_{me2}=2), (m_{me3}=4), and (m_{me4}=3).

[0185] 2. Feature Determination

[0186] 1. Determination of power drive crossover characteristics

[0187] The power drive crossover feature is determined based on the power drive crossover abnormal monitoring label and the number of power drive abnormal events. One possible determination method is to calculate a weighted average. For example, let the weight be the ratio of the number of abnormal events to the total number of abnormal events. For the power drive crossover abnormal monitoring labels (L_{pd1}), (L_{pd2}), (L_{pd3}), the total number of abnormal events (N_{pd}=n_{pd1}+n_{pd2}+n_{pd3}=5+3+4=12).

[0188] The power-driven crossover feature (F_{pd_cross}) is calculated as follows: (F_{pd_cross}=frac{n_{pd1}}{N_{pd}}L_{pd1}+frac{n_{pd2}}{N_{pd}}L_{pd2}+frac{n_{pd3}}{N_{pd}}L_{pd3}). For example, (L_{pd1}=0.6), (L_{pd2}=0.8), (L_{pd3}=0.7), then .

[0189] 2. Determination of motor unit cross-features

[0190] The motor unit crossing features are determined in a similar way. For the motor unit crossing abnormal monitoring labels (L_{me1}), (L_{me2}), (L_{me3}), (L_{me4}), the total number of abnormal events is (N_{me}=n_{me1}+n_{me2}+n_{me3}+n_{me4}=2+3+4+3=12).

[0191] The motor unit crossover feature (F_{me_cross}) is calculated as:

[0192] (F_{me_cross}=frac{n_{me1}}{N_{me}}L_{me1}+frac{n_{me2}}{N_{me}}L_{me2}+frac{n_{me3}}{N_{me}}L_{me3}+frac{n_{me4}}{N_{me}}L_{me4}).

[0193] For example, (L_{me1}=0.3), (L_{me2}=0.5), (L_{me3}=0.6), (L_{me4}=0.4),

[0194] but:

[0195] .

[0196] 3. Determination of power drive related characteristics

[0197] The power drive related features are determined according to the power drive related fault decision viewpoint and the number of power drive related abnormal events. The weighted average method is also used. For the power drive related fault decision viewpoints (D_{pd1}), (D_{pd2}), (D_{pd3}), the total number of related abnormal events (M_{pd}=m_{pd1}+m_{pd2}+m_{pd3}=4+3+5=12).

[0198] The power drive association feature (F_{pd_assoc}) is calculated as: (F_{pd_assoc}=frac{m_{pd1}}{M_{pd}}D_{pd1}+frac{m_{pd2}}{M_{pd}}D_{pd2}+frac{m_{pd3}}{M_{pd}}D_{pd3}). For example, (D_{pd1}=0.5), (D_{pd2}=0.7), (D_{pd3}=0.6), then .

[0199] 4. Determination of motor unit related features

[0200] For the motor unit associated fault decision viewpoints (D_{me1}), (D_{me2}), (D_{me3}), (D_{me3}), (D_{me4}), the total number of associated abnormal events is (M_{me}=m_{me1}+m_{me2}+m_{me3}+m_{me4}=3+2+4+3=12).

[0201] The motor unit association feature (F_{me_assoc}) is calculated as: (F_{me_assoc}=frac{m_{me1}}{M_{me}}D_{me1}+frac{m_{me2}}{M_{me}}D_{me2}+frac{m_{me3}}{M_{me}}D_{me3}+frac{m_{me4}}{M_{me}}D_{me4}). For example, (D_{me1}=0.4), (D_{me2}=0.3), (D_{me3}=0.5), (D_{me4}=0.4),

[0202] but:

[0203] .

[0204] 3. Fault Analysis Error Determination

[0205] 1. Error calculation method

[0206] The power drive cross-signal, the motor unit cross-signal, the power drive associated signature and the motor unit associated signature are subjected to fault analysis error determination to obtain a first fault analysis comparison error. A possible error calculation method is to use mean square error (MSE).

[0207] Assume (E_{1}) is the first fault analysis comparison error, then (E_{1}=frac{1}{4}[(F_{pd_cross}-F_{pd_assoc})^2+(F_{me_cross}-F_{me_assoc})^2+(F_{pd_assoc}-F_{pd_cross})^2+(F_{me_assoc}-F_{me_cross})^2]). Substituting the previously calculated values ​​into the equation, we get: (E_{1}=frac{1}{4}[(0.683-0.583)^2+(0.475-0.4)^2+(0.583-0.683)^2+(0.4-0.475)^2]=0.011).

[0208] It can be seen that by carefully classifying and characterizing the cross-abnormal monitoring labels and associated fault decision viewpoints related to power drive and motor units based on prior training annotations, the situation of different units in fault diagnosis can be analyzed more accurately. The characteristic values ​​are determined by weighted average and other methods, taking into account the influence of the number of abnormal events, making the features more representative. Finally, the first fault analysis comparison error is calculated by algorithms such as mean square error. This method based on quantitative comparison information can more accurately evaluate the difference between the fault diagnosis discriminant network and the associated fault decision algorithm, which helps to more effectively optimize the fault diagnosis discriminant network and improve the accuracy and reliability of the entire fault diagnosis system.

[0209] In some independent examples, when the abnormal monitoring causal restriction feature is the causal conduction information of the label heat map; the fault analysis error is determined for Y cross-abnormal monitoring labels according to the abnormal monitoring causal restriction feature to obtain the fault tracing error, including: determining the power drive abnormal monitoring discreteness according to the power drive cross-abnormal monitoring label and the number of power drive abnormal events, and determining the motor unit abnormal monitoring discreteness according to the motor unit cross-abnormal monitoring label and the number of motor unit abnormal events; determining the power drive associated fault discreteness according to the power drive associated fault decision-making viewpoint and the number of power drive associated abnormal events, and determining the motor unit associated fault discreteness according to the motor unit associated fault decision-making viewpoint and the number of motor unit associated abnormal events; performing fault analysis error determination on the power drive abnormal monitoring discreteness, the power drive cross feature, the motor unit abnormal monitoring discreteness, the motor unit cross feature, the power drive associated fault discreteness, the power drive associated feature, the motor unit associated fault discreteness and the motor unit associated feature to obtain the fault tracing error.

[0210] In other independent embodiments, the fault analysis error is determined for Y cross-abnormal monitoring labels and the scene fault characteristic values ​​according to the abnormal monitoring comparison restriction feature to obtain a second fault analysis comparison error, including: determining, according to the prior training annotations, from the Y cross-abnormal monitoring labels, the power drive cross-abnormal monitoring label corresponding to the power drive prior training annotation, the number of power drive abnormal events corresponding to the power drive cross-abnormal monitoring label, the motor unit cross-abnormal monitoring label corresponding to the motor unit prior training annotation, and the number of motor unit abnormal events corresponding to the motor unit cross-abnormal monitoring label; determining, according to the prior training annotations, from X scene fault characteristic values, the power drive fault characteristic value corresponding to the power drive prior training annotation, the power drive error frequency corresponding to the power drive fault characteristic value, the motor fault characteristic value corresponding to the motor unit prior training annotation, and the motor unit error frequency corresponding to the motor fault characteristic value; based on According to the power drive cross-abnormal monitoring label and the number of power drive abnormal events, the power drive cross-feature is determined, and according to the motor unit cross-abnormal monitoring label and the number of motor unit abnormal events, the motor unit cross-feature is determined; according to the power drive fault characteristic value and the power drive error frequency, the power drive stability coefficient is determined, and according to the motor fault characteristic value and the motor unit error frequency, the motor unit stability coefficient is determined; when the abnormal monitoring comparison restriction feature is the quantitative comparison information of the power drive label vector, the power drive cross-feature, the motor unit cross-feature and the power drive stability coefficient are subjected to fault analysis error determination to obtain the second fault analysis comparison error; when the abnormal monitoring comparison restriction feature is the quantitative comparison information of the motor unit label vector, the power drive cross-feature, the motor unit cross-feature and the motor unit stability coefficient are subjected to fault analysis error determination to obtain the second fault analysis comparison error.

[0211] Furthermore, the fault analysis error is determined for Y cross-abnormal monitoring labels and the prior cross-abnormal discrimination viewpoints according to the abnormal monitoring comparison restriction characteristics to obtain a third fault analysis comparison error, including: according to the prior training annotations, determining, from the Y cross-abnormal monitoring labels, the power drive cross-abnormal monitoring labels corresponding to the power drive prior training annotations, the number of power drive abnormal events corresponding to the power drive cross-abnormal monitoring labels, the motor unit cross-abnormal monitoring labels corresponding to the motor unit prior training annotations, and the number of motor unit abnormal events corresponding to the motor unit cross-abnormal monitoring labels; according to the prior training annotations, determining, from the X prior cross-abnormal discrimination viewpoints, the power drive prior cross-abnormal discrimination viewpoints corresponding to the power drive prior training annotations, the power drive reference abnormal event number corresponding to the power drive prior cross-abnormal discrimination viewpoints, the motor unit prior cross-abnormal discrimination viewpoints corresponding to the motor unit prior training annotations, and the number of motor unit reference abnormal events corresponding to the motor unit prior cross-abnormal discrimination viewpoints. ; According to the power drive cross-abnormal monitoring label and the number of power drive abnormal events, determine the power drive cross feature, and according to the motor unit cross-abnormal monitoring label and the number of motor unit abnormal events, determine the motor unit cross feature; According to the power drive prior cross-abnormal discrimination viewpoint and the number of power drive reference abnormal events, determine the power drive abnormal decision viewpoint, and according to the motor unit prior cross-abnormal discrimination viewpoint and the number of motor unit reference abnormal events, determine the motor unit abnormal decision viewpoint; When the abnormal monitoring comparison restriction feature is the quantized comparison information of the power drive label vector, the power drive cross feature, the motor unit cross feature and the power drive abnormal decision viewpoint are performed a fault analysis error determination to obtain the third fault analysis comparison error; When the abnormal monitoring comparison restriction feature is the quantized comparison information of the motor unit label vector, the power drive cross feature, the motor unit cross feature and the motor unit abnormal decision viewpoint are performed a fault analysis error determination to obtain the third fault analysis comparison error.

[0212] The above two independent embodiments are exemplarily introduced as follows.

[0213] 1. Determine the fault tracing error based on the causal transmission information of the label heat map

[0214] 1. Determination of Discreteness

[0215] 1. Power drive abnormal monitoring discreteness

[0216] The power drive abnormal monitoring discreteness is determined according to the power drive cross abnormal monitoring label and the number of power drive abnormal events. For example, the power drive cross abnormal monitoring label is a vector (vec{L}_{pd}=[l_{pd1}, l_{pd2}, cdots, l_{pdN}]) ((N) is the number of monitoring labels related to the power drive), and the number of power drive abnormal events is (n_{pd}).

[0217] One method of calculating the power drive abnormality monitor dispersion (D_{pd_monitor}) may be standard deviation calculation.

[0218] First calculate the mean (mu_{pd}=frac{1}{N}sum_{i=1}^{N}l_{pdi}), then determine:

[0219] (D_{pd_monitor}

[0220] =sqrt{frac{1}{N}sum_{i=1}^{N}(l_{pdi}-mu_{pd})^2}).

[0221] For example, (vec{L}_{pd}=[0.2, 0.3, 0.5, 0.4]), (n_{pd}=10), we can calculate (mu_{pd}=frac{0.2+0.3+0.5+0.4}{4}=0.35),

[0222] (D_{pd_monitor}=sqrt{frac{(0.2-0.35)^2+(0.3-0.35)^2+(0.5-0.35)^2+(0.4-0.35)^2}{4}}=0.1118).

[0223] 2. Abnormal monitoring discreteness of motor unit

[0224] For the motor unit, the motor unit abnormality monitoring discreteness (D_{me_monitor}) is also determined based on the motor unit cross-abnormality monitoring label (vec{L}_{me}=[l_{me1}, l_{me2}, cdots, l_{meM}]) ((M) is the number of monitoring labels related to the motor unit) and the number of motor unit abnormality events (n_{me}).

[0225] For example (vec{L}_{me}=[0.1, 0.2, 0.3, 0.4]), (n_{me}=8), calculate the average (mu_{me}=frac{0.1+0.2+0.3+0.4}{4}=0.25), (D_{me_monitor}=sqrt{frac{(0.1-0.25)^2+(0.2-0.25)^2+(0.3-0.25)^2+(0.4-0.25)^2}{4}}=0.1118).

[0226] 3. Power drive related fault dispersion

[0227] The power drive associated fault dispersion (D_{pd_assoc}) is determined according to the power drive associated fault decision viewpoint (vec{D}_{pd}=[d_{pd1}, d_{pd2}, cdots, d_{pdK}]) ((K) is the number related to the power drive associated fault decision) and the number of power drive associated abnormal events (m_{pd}).

[0228] For example, (vec{D}_{pd}=[0.3, 0.4, 0.5, 0.6]), (m_{pd}=12), calculate the mean (mu_{pd_assoc}=frac{0.3+0.4+0.5+0.6}{4}=0.45), (D_{pd_assoc}=sqrt{frac{(0.3-0.45)^2+(0.4-0.45)^2+(0.5-0.45)^2+(0.6-0.45)^2}{4}}=0.1118).

[0229] 4. Motor unit related fault discreteness

[0230] The motor unit associated fault discreteness (D_{me_assoc}) is determined according to the motor unit associated fault decision viewpoint (vec{D}_{me}=[d_{me1}, d_{me2}, cdots, d_{meL}]) ((L) is the quantity related to the motor unit associated fault decision) and the number of motor unit associated abnormal events (m_{me}).

[0231] For example, (vec{D}_{me}=[0.2, 0.3, 0.4, 0.5]), (m_{me}=10), calculate the mean (mu_{me_assoc}=frac{0.2+0.3+0.4+0.5}{4}=0.35), (D_{me_assoc}=sqrt{frac{(0.2-0.35)^2+(0.3-0.35)^2+(0.4-0.35)^2+(0.5-0.35)^2}{4}}=0.1118).

[0232] 2. Fault tracing error determination

[0233] 1. Fault analysis error calculation method

[0234] The fault analysis error of power drive abnormal monitoring discreteness, power drive cross characteristics, motor unit abnormal monitoring discreteness, motor unit cross characteristics, power drive associated fault discreteness, power drive associated characteristics, motor unit associated fault discreteness and motor unit associated characteristics is determined to obtain the fault tracing error.

[0235] For example, a comprehensive error calculation method is adopted, assuming that the fault tracing error is (E_{trace}), (E_{trace}=frac{1}{8}[(D_{pd_monitor}-F_{pd_cross})^2+(D_{me_monitor}-F_{me_cross})^2+(D_{pd_assoc}-F_{pd_assoc})^2+(D_{me_assoc}-F_{me_assoc})^2+(F_{pd_cross}-D_{pd_monitor})^2+(F_{me_cross}-D_{me_monitor})^2+(F_{pd_assoc}-D_{pd_assoc})^2+(F_{me_assoc}-D_{me_assoc})^2])

[0236] In the present invention, (F_{pd_cross}), (F_{me_cross}), (F_{pd_assoc}), (F_{me_assoc}) are the cross characteristics and association characteristics of the power drive and motor unit calculated previously (according to the calculation method in the previous section). Substituting the previously calculated values ​​into the calculation, the value of the fault tracing error (E_{trace}) can be obtained.

[0237] 2. Determine the comparison error of the second and third fault analysis based on the comparison information of different label vector quantization

[0238] 1. Determine the second fault analysis comparison error

[0239] 1. Determine the second fault analysis comparison error based on the power drive tag vector quantization comparison information

[0240] According to the prior training annotations, the power drive cross-abnormal monitoring labels corresponding to the power drive prior training annotations, the corresponding number of power drive abnormal events, the motor unit cross-abnormal monitoring labels corresponding to the motor unit prior training annotations, and the number of motor unit abnormal events are determined from the (Y) cross-abnormal monitoring labels. The power drive fault feature values ​​corresponding to the power drive prior training annotations, the corresponding power drive error frequency, the motor fault feature values ​​corresponding to the motor unit prior training annotations, and the motor unit error frequency are determined from the (X) scenario fault feature values.

[0241] For example, suppose (Y=10), in which the cross-abnormal monitoring tag related to power drive is determined, for example, the cross-abnormal monitoring tag of power drive is , the corresponding number of power drive abnormal events ; The motor unit cross abnormal monitoring label is , the corresponding number of abnormal events of the motor unit Assume (X=8), the power drive fault characteristic value is (f_{pd}=0.6), and the power drive error frequency is (r_{pd}=5); the motor fault characteristic value is (f_{me}=0.4), and the motor unit error frequency is (r_{me}=3).

[0242] Determine the power drive crossover feature based on the power drive crossover anomaly monitoring tag and the number of power drive anomaly events (Calculation method is the same as in the previous section), determine the motor unit crossover feature based on the motor unit crossover abnormality monitoring label and the number of motor unit abnormal events The power drive stability factor (S_{pd}) is determined based on the power drive fault characteristic value and the power drive error frequency, for example , the motor unit stability factor (S_{me}) is determined according to the motor fault characteristic value and the motor unit error frequency.

[0243] When the abnormal monitoring comparison restriction feature is the quantitative comparison information of the power drive label vector, the fault analysis error is determined for the power drive cross feature, the motor unit cross feature and the power drive stability coefficient to obtain the second fault analysis comparison error (E_{2}). For example, the mean square error calculation method can be used. .

[0244] 2. Determine the second fault analysis comparison error based on the motor unit label vector quantization comparison information

[0245] Similarly, when the abnormal monitoring comparison restriction feature is the quantized comparison information of the motor unit label vector, the fault analysis error is determined for the power drive crossover feature, the motor unit crossover feature, and the motor unit stability coefficient to obtain the second fault analysis comparison error (E_{2}). Using a similar mean square error calculation method, .

[0246] (II) Determine the third fault analysis comparison error

[0247] 1. Determine the third fault analysis comparison error based on the power drive tag vector quantization comparison information

[0248] According to the prior training annotations, the power drive cross-abnormal monitoring labels corresponding to the power drive prior training annotations, the corresponding number of power drive abnormal events, the motor unit cross-abnormal monitoring labels corresponding to the motor unit prior training annotations, and the number of motor unit abnormal events are determined from the (Y) cross-abnormal monitoring labels. The power drive prior cross-abnormal discrimination viewpoints corresponding to the power drive prior training annotations, the corresponding number of power drive reference abnormal events, the motor unit prior cross-abnormal discrimination viewpoints corresponding to the motor unit prior training annotations, and the number of motor unit reference abnormal events are determined from the (X) prior cross-abnormal discrimination viewpoints.

[0249] For example, suppose (Y=10), determine that the power drive cross-abnormal monitoring label is (L_{pd}^{''}=[l_{pd1}^{''}, l_{pd2}^{''}]), and the corresponding number of power drive abnormal events is (n_{pd}^{''}=3); the motor unit cross-abnormal monitoring label is (L_{me}^{''}=[l_{me1}^{''}]), and the corresponding number of motor unit abnormal events is (n_{me}^{''}=2). Assume (X=8), the power drive prior cross-abnormal discrimination viewpoint is (D_{pd}^{''}=[d_{pd1}^{''}, d_{pd2}^{''}]), and the corresponding number of power drive reference abnormal events (m_{pd}^{''}=4); the motor unit prior cross-abnormal discrimination viewpoint is (D_{me}^{''}=[d_{me1}^{''}]), and the corresponding number of motor unit reference abnormal events (m_{me}^{''}=3).

[0250] The power drive cross feature (F_{pd_cross}^{''}) is determined based on the power drive cross abnormality monitoring label and the number of power drive abnormal events, and the motor unit cross feature (F_{me_cross}^{''}) is determined based on the motor unit cross abnormality monitoring label and the number of motor unit abnormal events. The power drive abnormality decision view (O_{pd}) is determined based on the power drive prior cross abnormality discrimination view and the number of power drive reference abnormal events (for example, weighted average calculation can be used), and the motor unit abnormality decision view (O_{me}) is determined based on the motor unit prior cross abnormality discrimination view and the number of motor unit reference abnormal events.

[0251] When the abnormal monitoring comparison restriction feature is the quantitative comparison information of the power drive label vector, the fault analysis error of the power drive cross feature, the motor unit cross feature and the power drive abnormal decision viewpoint is determined to obtain the third fault analysis comparison error (E_{3}). For example, the mean square error calculation method is adopted, (E_{3}=frac{1}{3}[(F_{pd_cross}^{''}-O_{pd})^2+(F_{me_cross}^{''}-O_{pd})^2+(O_{pd}-F_{pd_cross}^{''})^2]).

[0252] 2. Determine the third fault analysis comparison error based on the motor unit label vector quantization comparison information

[0253] When the abnormal monitoring comparison restriction feature is the quantitative comparison information of the motor unit label vector, the fault analysis error of the power drive cross feature, the motor unit cross feature and the motor unit abnormal decision viewpoint is determined to obtain the third fault analysis comparison error (E_{3}). A similar mean square error calculation method is used, (E_{3}=frac{1}{3}[(F_{pd_cross}^{''}-O_{me})^2+(F_{me_cross}^{''}-O_{me})^2+(O_{me}-F_{pd_cross}^{''})^2]).

[0254] It can be seen that the technical effect of the above-mentioned embodiment is to more accurately determine different types of fault analysis errors through a variety of specific numerical calculations and algorithm examples. When determining the fault tracing error, based on the causal conduction information of the label heat map, by considering multiple factors such as the monitoring discreteness, cross-features and correlation features of different units, it is possible to more deeply explore the errors related to the root cause of the fault. For the second and third fault analysis comparison errors, according to different label vector quantization comparison information, detailed analysis is performed on the parameters related to the power drive and motor unit respectively, which helps to accurately evaluate the differences between the fault diagnosis discriminant network and other related factors (such as scene fault feature values, prior cross-abnormal discrimination views), thereby providing a more accurate basis for optimizing the fault diagnosis discriminant network and improving the accuracy and reliability of the entire fault diagnosis system.

[0255] In summary, firstly, by mining the light and heavy load state vectors of the embedded chip operation records, the complex operation records can be converted into the target light and heavy load state vector chain. This method effectively integrates the key information of the system under different load states and lays a comprehensive data foundation for subsequent fault analysis. Then, the target cross-jump state vector chain is refined according to the trend change threshold eigenvalue, and the vector chain with special state changes is accurately screened out, which helps to focus on the key state where the fault may exist. The initial cross-abnormal monitoring label is generated by the fault diagnosis and discrimination network, and the fault situation is preliminarily judged with its powerful data analysis ability, which improves the efficiency of fault diagnosis. Then, the target cross-abnormal monitoring label is refined according to the cross trend change threshold eigenvalue, which further eliminates the misjudgment caused by short-term interference and greatly improves the accuracy of fault diagnosis. Finally, the target fault diagnosis event is determined and reported to the host computer system, so that the host computer can take corresponding measures in time to ensure the stable operation of the electromechanical servo system, reduce the loss caused by the failure to handle the fault in time, and improve the reliability and safety of the entire electromechanical servo system.

[0256] Further, Figure 2 FIG. 2 is a schematic diagram of a structure of a fault monitoring server 200 provided by the present invention. Figure 2 The fault monitoring server 200 shown includes a processor 210, and the processor 210 can call and run a computer program from a memory to implement the method of the present invention.

[0257] Alternatively, if Figure 2 As shown, the fault monitoring server 200 may further include a memory 230. The processor 210 may call and run a computer program from the memory 230 to implement the method of the present invention.

[0258] The memory 230 may be a separate device independent of the processor 210 , or may be integrated into the processor 210 .

[0259] Alternatively, if Figure 2 As shown, the fault monitoring server 200 may further include a transceiver 220, and the processor 210 may control the transceiver 220 to interact with other devices, specifically, may send information or data to other devices, or receive information or data sent by other devices.

[0260] Optionally, the fault monitoring server 200 may implement the corresponding processes corresponding to the storage engine or components in the storage engine (such as a processing module) or a device deployed with a storage engine in each method of the present invention, which will not be described in detail here for the sake of brevity.

[0261] It should be understood that the processor of the present invention may be an integrated circuit chip having signal processing capabilities.

[0262] It is understood that the memory in the present invention can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memory. It should be noted that the memory of the systems and methods described herein is intended to include but is not limited to suitable types of memory.

[0263] Based on the above, a readable storage medium is provided, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the steps of the above method are implemented.

[0264] It should be noted that, in this article, the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the method and device in the embodiment of the present invention is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0265] It should be understood that the formula parameter content recorded in the above technical solution is in LaTeX format (an existing formula editing format), and those skilled in the art can restore the corresponding original formula expression without any doubt based on the formula parameter content recorded in the above technical solution. For example, " " is the multiplication operator symbol (dot product), "vec" means vector, etc.

[0266] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, disk, CD), including several instructions for a terminal (which can be a mobile phone, computer, server, or network equipment, etc.) to execute the methods described in each embodiment of the present invention. The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are only illustrative and not restrictive. Under the enlightenment of the present invention, those skilled in the art can also make many forms without departing from the purpose of the present invention and the scope of protection of the present invention, all of which belong to the protection of the present invention.

Claims

1. A fault monitoring method applied to an electromechanical servo system, characterized in that: The method is implemented by a fault monitoring server, and the method includes: Mining the light and heavy load state vectors of the X embedded chip operation records to be analyzed corresponding to the X initially screened suspected fault events, respectively, to obtain a target light and heavy load state vector chain corresponding to each embedded chip operation record to be analyzed; According to the trend change threshold characteristic value, Y target cross jump state vector chains are extracted from the X target light and heavy load state vector chains, wherein Y is an integer not less than 1 and not greater than X; The Y target cross-jump state vector chains are respectively transmitted to the fault diagnosis and discrimination network, and Y initial cross-abnormal monitoring labels are generated through the fault diagnosis and discrimination network; According to the cross trend change threshold characteristic value, Z target cross anomaly monitoring labels are extracted from Y initial cross anomaly monitoring labels, wherein Z is an integer not less than 1 and not greater than Y; The initially screened suspected fault events corresponding to the Z target cross-abnormal monitoring tags are determined as Z target fault diagnosis events, and the Z target fault diagnosis events are reported to the host computer system.

2. The method according to claim 1, characterized in that The X embedded chip operation records to be analyzed belong to the same target electromechanical servo system; the light and heavy load state vector mining is used to perform digital discretization processing on the continuous sensor monitoring signals corresponding to the X embedded chip operation records to be analyzed; the fault unit corresponding to the target fault diagnosis event includes at least one of a power drive unit, a motor unit and a reducer unit.

3. The method according to any one of claims 1 to 2, characterized in that: The optimization method of the fault diagnosis and discrimination network comprises: Performing light and heavy load state vector mining on each of the X embedded chip operation record samples to obtain a light and heavy load state vector chain sample corresponding to each of the embedded chip operation record samples, wherein each of the light and heavy load state vector chain samples corresponds to a priori training annotation, and X is an integer not less than 1; According to the trend change threshold characteristic value, Y cross-jump state vector chain samples are extracted from the X light-load and heavy-load state vector chain samples, wherein Y is an integer not less than 1 and not greater than X; The Y cross-jump state vector chain samples are respectively transmitted to the fault diagnosis and discrimination network, and Y cross-abnormal monitoring labels are generated by the fault diagnosis and discrimination network; The Y cross-jump state vector chain samples are respectively input into the associated fault decision algorithm, and Y associated fault decision viewpoints are generated by the associated fault decision algorithm; According to the abnormal monitoring comparison restriction feature, the fault analysis error is determined for the Y cross-abnormal monitoring tags and the Y associated fault decision viewpoints to obtain a first fault analysis comparison error; According to the first fault analysis comparison error, the neural network weights of the fault diagnosis and discrimination network are optimized.

4. The method according to claim 3, characterized in that After determining the fault analysis error of the Y cross-abnormal monitoring tags and the Y associated fault decision viewpoints according to the abnormal monitoring comparison restriction feature and obtaining the first fault analysis comparison error, the method further includes: According to the abnormal monitoring causal restriction characteristics, the fault analysis error is determined for the Y cross abnormal monitoring tags to obtain the fault tracing error; The optimizing the neural network weights of the fault diagnosis and discrimination network according to the first fault analysis comparison error includes: The neural network weights of the fault diagnosis and discrimination network are optimized according to the first fault analysis comparison error and the fault tracing error.

5. The method according to claim 3, characterized in that The extracting Y cross-jump state vector chain samples from the X light-load and heavy-load state vector chain samples according to the trend change threshold characteristic value includes: According to the prior training annotations, a scenario fault diagnosis network is determined, and the light and heavy load state vector chain samples corresponding to the prior training annotations are passed into the scenario fault diagnosis network, and the scenario fault feature values ​​corresponding to the light and heavy load state vector chain samples are generated through the scenario fault diagnosis network; From the X light and heavy load state vector chain samples, determine the light and heavy load state vector chain samples whose scenario fault characteristic value is not less than the trend change threshold characteristic value, so as to obtain Y cross jump state vector chain samples; After the Y cross-jump state vector chain samples are respectively transmitted to the fault diagnosis and discrimination network and Y cross-abnormal monitoring labels are generated by the fault diagnosis and discrimination network, the method further includes: According to the abnormal monitoring comparison restriction feature, the fault analysis error is determined for the Y cross-abnormal monitoring tags and the scene fault feature value to obtain a second fault analysis comparison error; According to the second fault analysis comparison error, the neural network weights of the fault diagnosis and discrimination network are optimized.

6. The method according to claim 3, characterized in that After the Y cross-jump state vector chain samples are respectively transmitted to the fault diagnosis and discrimination network and Y cross-abnormal monitoring labels are generated by the fault diagnosis and discrimination network, the method further includes: Obtaining a priori cross-abnormal discrimination opinions corresponding to each of the embedded chip operation record samples; According to the abnormal monitoring comparison restriction feature, the fault analysis error is determined for the Y cross-abnormal monitoring labels and the prior cross-abnormal discrimination viewpoint to obtain a third fault analysis comparison error; According to the third fault analysis comparison error, the neural network weights of the fault diagnosis and discrimination network are optimized.

7. The method according to claim 5, characterized in that The scenario fault diagnosis network includes a power drive unit fault diagnosis subnet and a motor unit fault diagnosis subnet, and the scenario fault characteristic value includes a power drive fault characteristic value and a motor fault characteristic value; The method of determining a scenario fault diagnosis network according to the prior training annotations, and transferring the light and heavy load state vector chain samples corresponding to the prior training annotations to the scenario fault diagnosis network, and generating scenario fault feature values ​​corresponding to the light and heavy load state vector chain samples through the scenario fault diagnosis network, includes: When the prior training annotation is a power drive fault annotation, the power drive unit fault diagnosis subnet is determined, and the light and heavy load state vector chain samples corresponding to the power drive fault annotation are transmitted to the power drive unit fault diagnosis subnet, and the power drive fault feature values ​​corresponding to the light and heavy load state vector chain samples are generated by the power drive unit fault diagnosis subnet; When the prior training annotation is a motor fault annotation, the motor unit fault diagnosis subnet is determined, and the light and heavy load state vector chain samples corresponding to the motor fault annotation are transmitted to the motor unit fault diagnosis subnet, and the motor fault feature values ​​corresponding to the light and heavy load state vector chain samples are generated by the motor unit fault diagnosis subnet; The step of determining, from the X light and heavy load state vector chain samples, a light and heavy load state vector chain sample whose scenario fault characteristic value is not less than the trend change threshold characteristic value to obtain Y cross-jump state vector chain samples comprises: From the X light and heavy load state vector chain samples, determine the light and heavy load state vector chain samples where the power drive fault characteristic value and the motor fault characteristic value are not less than the trend change threshold characteristic value, so as to obtain Y cross jump state vector chain samples; Wherein, when the prior training annotation is a motor fault annotation, the motor unit fault diagnosis subnet is determined, and the light and heavy load state vector chain samples corresponding to the motor fault annotation are transmitted to the motor unit fault diagnosis subnet, and after the motor fault feature values ​​corresponding to the light and heavy load state vector chain samples are generated by the motor unit fault diagnosis subnet, the method further includes: Determining a power drive fault analysis error according to the power drive fault annotation and the power drive fault characteristic value, and optimizing a neural network weight of the power drive unit fault diagnosis subnet according to the power drive fault analysis error; The motor fault analysis error is determined according to the motor fault annotation and the motor fault characteristic value, and the neural network weight of the motor unit fault diagnosis subnet is optimized according to the motor fault analysis error.

8. The method according to any one of claims 4 to 5, characterized in that The light-heavy load state vector mining is performed on each of the X embedded chip operation record samples to obtain a light-heavy load state vector chain sample corresponding to each of the embedded chip operation record samples, including: Obtaining a platform interaction description vector of an electromechanical servo system sample corresponding to the embedded chip operation record sample, wherein the platform interaction description vector is used to characterize a servo system state element between each electromechanical servo system sample and a suspected fault event sample; According to the electromechanical servo system sample and the platform interaction description vector, the light and heavy load state vectors of the embedded chip operation record samples are mined to obtain the light and heavy load state vector chain samples corresponding to each of the embedded chip operation record samples.

9. The method according to claim 4, characterized in that When the abnormal monitoring comparison restriction feature is the quantized comparison information of the label vector; the fault analysis error is determined for the Y cross-abnormal monitoring labels and the Y associated fault decision viewpoints according to the abnormal monitoring comparison restriction feature to obtain a first fault analysis comparison error, including: According to the prior training annotations, determining, from the Y cross-abnormal monitoring labels, a power drive cross-abnormal monitoring label corresponding to the power drive prior training annotation, a number of power drive abnormal events corresponding to the power drive cross-abnormal monitoring label, a motor unit cross-abnormal monitoring label corresponding to the motor unit prior training annotation, and a number of motor unit abnormal events corresponding to the motor unit cross-abnormal monitoring label; According to the prior training annotations, determining, from the Y associated fault decision viewpoints, a power drive associated fault decision viewpoint corresponding to the power drive prior training annotations, a number of power drive associated abnormal events corresponding to the power drive associated fault decision viewpoints, a motor unit associated fault decision viewpoint corresponding to the motor unit prior training annotations, and a number of motor unit associated abnormal events corresponding to the motor unit associated fault decision viewpoints; Determine a power drive cross feature according to the power drive cross abnormality monitoring tag and the number of power drive abnormal events, and determine a motor unit cross feature according to the motor unit cross abnormality monitoring tag and the number of motor unit abnormal events; Determine a power drive associated feature according to the power drive associated fault decision viewpoint and the number of power drive associated abnormal events, and determine a motor unit associated feature according to the motor unit associated fault decision viewpoint and the number of motor unit associated abnormal events; A fault analysis error is determined on the power drive cross-feature, the motor unit cross-feature, the power drive associated feature, and the motor unit associated feature to obtain the first fault analysis comparison error.

10. A fault monitoring server, characterized in that: The method comprises at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the method according to any one of claims 1 to 9.

Images