Device Fault Diagnosis Method and Device Based on Sensor Data

Abstract

The present invention provides a device fault diagnosis method and device based on sensor data, which relates to the technical field of data processing. After performing chaotic randomization feature extraction on the operation state data of the device to be measured through a pre-trained neural network, the fault diagnosis model is used for classification and identification to determine the device fault diagnosis result. Chaotic randomization feature extraction helps to break the limitation of local optimal solutions and increase the possibility of exploring different solutions. In addition, the neural network used for chaotic randomization feature extraction is constructed based on self-organizing feedback, and the neurons in the neural network cooperate effectively, which can help the network better learn the internal structure in the data. It can not only capture the dynamic changes and non-linear relationships in the data more accurately, but also make full use of the distribution characteristics of the feature space of the data to improve the effect of fault identification.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular, to a device fault diagnosis method and device based on sensor data. Background Art

[0002] With the wide application of industrial devices, the monitoring of the operating state and fault diagnosis of devices have become important means to ensure the safety and reliability of industrial systems. Faults in industrial devices can not only lead to production interruptions but also may cause safety accidents. Therefore, an efficient and accurate fault diagnosis method is needed to reduce potential risks.

[0003] Among them, there are many challenges in data acquisition, processing, and analysis in device fault diagnosis. The operating environment of the device is complex and changeable, and the data characteristics change over time. However, existing feature extraction methods often cannot reflect these changes in real time or effectively, and it is difficult to effectively capture the changes in the data, resulting in the feature extraction results not being able to accurately represent the actual operating conditions of the device. In addition, the data generated by the device is usually multi-dimensional and interrelated. When existing technologies perform feature extraction, they often process each feature separately, ignoring the internal connections between the features. This may lead to incomplete feature information extraction, affecting the accuracy of subsequent analysis and fault diagnosis.

[0004] In addition, in the process of high-dimensional data classification, effective feature space analysis can help identify data patterns and abnormal situations. However, existing methods usually only focus on reducing dimensions or simply applying some general preprocessing steps, failing to deeply explore the structural information of the feature space and not fully utilizing the distribution characteristics of the feature space, resulting in insufficient model generalization ability and poor fault recognition effect for unknown device states. Summary of the Invention

[0005] To solve the above technical problems, embodiments of the present invention provide a device fault diagnosis method and device based on sensor data, which can more accurately capture the dynamic changes and non-linear relationships in the data, and fully utilize the distribution characteristics of the feature space of the data to improve the effect of fault recognition.

[0006] In a first aspect, embodiments of the present invention provide a device fault diagnosis method based on sensor data. The method includes: obtaining the operating state data of a device to be measured; using a pre-trained neural network to perform chaotic randomization feature extraction on the operating state data to determine the target features of the operating state data; wherein the neural network is constructed based on self-organizing feedback; inputting the target features into a preset fault diagnosis model to determine the fault classification result corresponding to the target features; and determining the fault state of the device to be measured based on the fault classification result.

[0007] Combined with the first aspect, the embodiment of the present invention provides the first implementation manner of the first aspect. Among them, the steps of using a pre-trained neural network to extract chaotic randomization features from the operation state data and determining the target features of the operation state data include: inputting the operation state data into a preset neural network, and based on the oscillation state of the neurons of the neural network, performing chaotic randomization search on the operation state data to determine the neuron output corresponding to each neuron of the neural network; determining the self-organizing feedback amount corresponding to the neural network according to the neuron output of each neuron; adjusting the neuron output according to the self-organizing feedback amount to determine the target features of the operation state data.

[0008] Combined with the first aspect, the embodiment of the present invention provides the second implementation manner of the first aspect. Among them, the steps of inputting the operation state data into a preset neural network, and based on the oscillation state of the neurons of the neural network, performing chaotic randomization search on the operation state data to determine the neuron output corresponding to each neuron of the neural network include: obtaining the oscillation parameters of the neurons of the neural network, and based on the oscillation parameters, performing chaotic randomization search on the operation state data to determine the neuron output corresponding to the neural network.

[0009] Combined with the first aspect, the embodiment of the present invention provides the third implementation manner of the first aspect. Among them, the calculation method of the oscillation parameters of the neurons includes: using a preset chaotic function to perform chaotic random initialization on the initial weights and biases of the neural network; where the chaotic function includes a preset oscillation phase and angle parameters; based on the oscillation phase and a preset oscillation control parameter, setting the initial oscillation frequency and amplitude of the neurons of the neural network; inputting a preset training sample set into the neural network to adaptively adjust the initial oscillation frequency and amplitude; and iterating the preset oscillation phase and angle parameters based on the chaotic perturbation term corresponding to the oscillation phase; until the neural network meets the preset training requirements to obtain the final oscillation parameters of the neurons.

[0010] Combined with the first aspect, the embodiment of the present invention provides the fourth implementation manner of the first aspect. Among them, the steps of adjusting the neuron output according to the self-organizing feedback amount to determine the target features of the operation state data include: updating the weight parameters of the neural network based on the self-organizing feedback amount to adjust the neuron output.

[0011] Combined with the first aspect, an embodiment of the present invention provides a fifth implementation manner of the first aspect. Among them, the method for constructing a preset training sample set includes: obtaining pre-collected device operation monitoring data; performing data annotation on the device operation monitoring data according to the device operation state indicated by the device operation monitoring data to construct an initial sample set; determining the feature distribution of the initial sample set; based on the feature distribution, determining the target position to be augmented in the initial sample set; performing data augmentation on the target position to be augmented through a preset sample augmentation algorithm to obtain augmented samples; and combining the augmented samples and the initial sample set to construct a training sample set.

[0012] Combined with the first aspect, an embodiment of the present invention provides a sixth implementation manner of the first aspect. Among them, the step of determining the target augmentation position in the initial sample set based on the feature distribution includes: determining the sample augmentation weight of the minority class samples corresponding to the initial sample set based on the feature distribution; and determining the target position to be augmented corresponding to each sample according to the sample augmentation weight and the Euclidean distance between each sample of the initial sample set.

[0013] Combined with the first aspect, an embodiment of the present invention provides a seventh implementation manner of the first aspect. Among them, the step of performing data augmentation on the target position to be augmented through a preset sample augmentation algorithm to obtain augmented samples includes: performing data interpolation on the position to be augmented based on the sample augmentation weight to generate the augmented samples corresponding to the current sample.

[0014] Combined with the first aspect, an embodiment of the present invention provides an eighth implementation manner of the first aspect. Among them, the step of determining the feature distribution of the initial sample set includes: determining the feature mean of the initial sample set, and calculating the feature standard deviation of the initial sample set based on the feature mean; and performing feature space normalization processing on the initial sample set based on the feature standard deviation to determine the feature distribution of the initial sample set.

[0015] In the second aspect, an embodiment of the present invention provides a device for diagnosing device faults based on sensor data. The device includes: a data acquisition module for acquiring the operation state data of the device to be measured; a data processing module for extracting chaotic randomization features from the operation state data by using a pre-trained neural network to determine the target features of the operation state data, where the neural network is constructed based on self-organization feedback; an execution module for inputting the target features into a preset fault diagnosis model to determine the fault classification result corresponding to the target features; and an output module for determining the fault state of the device to be measured based on the fault classification result.

[0016] The embodiments of the present invention bring the following beneficial effects: A device fault diagnosis method and device based on sensor data provided by the embodiments of the present invention can break through the limitation of local optimal solutions, effectively identify the internal structure of data, more accurately capture the dynamic changes and non-linear relationships in the data, and thus obtain more accurate target features. Therefore, it can make full use of the distribution characteristics of the feature space of the data, adapt to the characteristics of different data sets or changes in the device state for fault identification, and improve the effect of fault identification.

[0017] Other features and advantages of the present invention will be described in the following specification, and, in part, will be obvious from the specification, or will be understood by implementing the present invention. The objectives and other advantages of the present invention are achieved and obtained by the structures specifically pointed out in the specification, claims, and drawings.

[0018] To make the above objectives, features, and advantages of the present invention more obvious and understandable, the following specifically enumerates preferred embodiments and, in conjunction with the accompanying drawings, makes a detailed description as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0020] Figure 1 It is a flowchart of a device fault diagnosis method based on sensor data provided by an embodiment of the present invention;

[0021] Figure 2 It is a flowchart of another device fault diagnosis method based on sensor data provided by an embodiment of the present invention;

[0022] Figure 3 It is a schematic diagram showing the influence of different initialization methods on the training loss provided by an embodiment of the present invention;

[0023] Figure 4 It is a schematic diagram showing the performance comparison of different algorithms under noise data provided by an embodiment of the present invention;

[0024] Figure 5 It is a flowchart of a method for constructing a training sample set provided by an embodiment of the present invention;

[0025] Figure 6 It is a schematic structural diagram of a device fault diagnosis device based on sensor data provided by an embodiment of the present invention;

[0026] Figure 7A schematic structural diagram of an electronic device provided by an embodiment of the present invention. Detailed implementation manners

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are some, rather than all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0028] The embodiments of the present invention provide a device fault diagnosis method and device based on sensor data, which can more accurately capture the dynamic changes and non-linear relationships in the data, and make full use of the distribution characteristics of the feature space of the data to improve the effect of fault identification.

[0029] To facilitate the understanding of this embodiment, first, a device fault diagnosis method based on sensor data disclosed in the embodiments of the present invention will be introduced in detail. Refer to Figure 1 The flowchart of a device fault diagnosis method based on sensor data shown, and this method may include the following steps:

[0030] Step S102, obtain the operation status data of the device to be measured.

[0031] The data collection source of the present invention is sensors installed on various industrial devices, including but not limited to temperature sensors, vibration sensors, and pressure sensors. These sensors can monitor the operation status of the device in real time and collect key performance parameters. Among them, sensors can be installed at key parts of the device according to the operation characteristics and monitoring requirements of the device. For example, temperature sensors are installed near components that may generate heat; vibration sensors are fixed on the mechanical main structure to monitor the vibration level during device operation; pressure sensors are installed on pipes where liquid or gas pressure needs to be monitored; current and voltage sensors are connected to the power supply line. The sensors continuously monitor various parameters of the device and convert the collected device data into electrical signals. These electrical signals are converted into digital signals through an analog-to-digital converter, and then are preliminarily processed in the micro-control unit of the sensor, such as filtering, amplification, and device data formatting. Further, the processed device data can be transmitted to a central database or cloud platform through Internet of Things technology. The wireless transmission technologies that can be used include: Wi-Fi: suitable for the interior of factories or warehouses; cellular networks: suitable for devices with dispersed geographical locations; LPWAN technologies (such as LoRa, Sigfox): suitable for long-distance and low-power transmission; Bluetooth: suitable for short-distance communication.

[0032] In one embodiment, the attributes of the device data include: temperature, specifically referring to the ambient temperature during device operation, in degrees Celsius; vibration frequency, specifically referring to the vibration frequency during device operation, in Hertz; pressure, specifically referring to the pressure value during device operation, in Pascals; current, specifically referring to the amount of current used by the device, in Amperes; voltage, specifically referring to the voltage value used by the device, in Volts; operating duration, specifically referring to the operating time of the device from startup to the data acquisition moment, in hours; fault code, specifically referring to the error code recorded when the device fails; maintenance record, specifically referring to the time of the last maintenance of the device; production batch, specifically referring to the production batch number of the device; model, specifically referring to the model identification of the device.

[0033] It should be noted that this embodiment is only to illustrate a device data format and type of the present invention. In actual applications, the attributes of the data are usually more than 10 attributes, and the number of data attributes may reach dozens or even hundreds.

[0034] Step S104: Use a pre-trained neural network to perform chaotic randomization feature extraction on the operating state data to determine the target features of the operating state data.

[0035] Among them, in the embodiment of the present invention, by introducing chaos initialization to the neural network, a chaotic sequence (such as Logistic mapping, Henon mapping, etc.) is used to initialize the weights or parameters of the neural network. The chaotic sequence is highly sensitive to the initial conditions and has pseudo-random properties. This characteristic helps to break through the limitation of the local optimal solution and increase the possibility of exploring different solutions. In addition, using a neural network constructed by self-organizing feedback for feature extraction, the effective cooperation between the neurons of the neural network can help the network better learn the internal structure in the data and can more accurately capture the dynamic changes and non-linear relationships in the data.

[0036] Step S106: Input the target features into a preset fault diagnosis model to determine the fault classification result corresponding to the target features.

[0037] The embodiment of the present invention uses a preset fault diagnosis model for fault diagnosis and classification. Among them, the fault diagnosis model adopted in the embodiment of the present invention.

[0038] Step S108: Determine the fault state of the device to be tested based on the fault classification result.

[0039] Through the fault diagnosis model, classifying and identifying the target features, outputting the probability distribution belonging to each category, and determining the category with a higher probability from them to obtain the device fault state classification result indicated by the target features, among which, including "normal operation" and "fault state", a total of 2 categories.

[0040] An equipment fault diagnosis method based on sensor data provided by an embodiment of the present invention can break through the limitation of local optimal solutions, effectively identify the internal structure of data, more accurately capture the dynamic changes and non-linear relationships in the data, and thus obtain more accurate target features. It can also make full use of the distribution characteristics of the feature space of the data, adapt to the characteristics of different data sets or changes in the equipment state for fault identification, and improve the effect of fault identification.

[0041] Further, an embodiment of the present invention also provides another equipment fault diagnosis method based on sensor data to illustrate the above step S104. Figure 2 The flowchart of the embodiment of the present invention is shown. Referring to Figure 2 , it includes the following steps:

[0042] Step S202, obtain the operation state data of the device to be measured.

[0043] Step S204, input the operation state data into a preset neural network, and based on the oscillation state of the neurons of the neural network, perform chaotic random search on the operation state data to determine the neuron output corresponding to each neuron of the neural network.

[0044] In the prior art, the output of a neuron is usually based on the linear weighted result of the input data, and the final value is output after being processed by an activation function. Such as traditional models like multi-layer perceptrons, convolutional neural networks, etc. Among them, when a neural network processes complex problems such as equipment fault diagnosis, the neurons have self-dynamic adjustment and non-linear perturbations. The behavior of these neurons in the prior art mainly depends on the weighted sum of the inputs, and does not fully consider the dynamic influence of the corresponding "oscillation state" of the neurons on the output of the neural network, resulting in a weak global search ability of the network, being difficult to fully capture the dynamic changes and non-linear features of the input data, and being prone to falling into local optima, resulting in the extracted features lacking the ability to represent the abnormal behavior of the equipment.

[0045] In this regard, the embodiment of the present invention determines the final target feature based on the self-organizing feedback corresponding to the neurons of the neural network. The target feature is determined through the collaborative optimization among the neurons inside the neural network, which can enhance the global exploration ability, accelerate the convergence speed, more effectively jump out of local extrema, and approximate the global optimal solution in a shorter time. Further, the embodiment of the present invention simulates the nonlinear dynamic behavior in the chaotic system, sets preset oscillation parameters (such as oscillation frequency and amplitude) for the neurons, and the changes in the oscillation frequency and amplitude simulate the nonlinear characteristics of many systems in nature, such as climate change, ecological systems, etc. These systems exhibit high dynamics and complexity under certain conditions, with irregular and unpredictable characteristics. By introducing dynamic changes similar to chaotic oscillations, the neural network can enhance the exploration ability during the training process, avoid falling into local optimal solutions, and thus improve the search ability. The chaotic oscillations make the state of the neurons not only depend on the linear weighted sum of the input signals, but also be affected by their internal nonlinear dynamic behavior, which is combined with the oscillation state of the current neuron to form a dynamic correction to the output. By setting dynamic oscillation frequency and amplitude, each neuron of the neural network can generate nonlinear perturbations during the training process, which helps to capture complex patterns in the input data. Moreover, the real-time changes of the features (such as environmental temperature fluctuations or abnormal vibration frequencies) can be dynamically coupled with the chaotic oscillation state of the neurons, and the neurons can effectively cooperate, making full use of the internal correlation between device features to analyze complex fault patterns and improving the search ability of the neural network when extracting features.

[0046] Specifically, the oscillation parameters of the neurons of the neural network are set through the following steps:

[0047] 1) Use a preset chaotic function to perform chaotic random initialization on the initial weights and biases of the neural network.

[0048] To prevent the appearance of symmetric solutions and provide a larger search space for the neural network, first, the parameters of the neural network are initialized by performing chaotic randomization on the initial weights and biases of the neural network. The embodiment of the present invention uses a chaotic function (such as a sine function or a cosine function) to perturb the initial weights, so that the initialized weights and biases of each neuron have a certain nonlinear deviation. Based on this, even the originally symmetric weights will have their symmetry broken by chaotic perturbations. Specifically, the chaotic random initialization of the initial weights and biases of the neural network is expressed as:

[0049]

[0050]

[0051] In the formula, is the initial weight of the neural network, is the initial bias of the neural network, is the oscillation phase, is the first initialization angle parameter, is the second initialization angle parameter, is the fourth initialization angle parameter, is the constant offset, is the small perturbation term. Preferably, = 0.5, = 0.5, = 0.1, = 0.01, , .

[0052] Among them, the first initialization angle parameter is used to control the oscillation phase of neurons, prevent the initial state of neurons from being too symmetric, and enhance the exploration ability of the network. During the training process of the neural network, each neuron oscillates according to this initial phase, affecting the activation state of the neuron. If the initial phases of all neurons are the same, it may lead to symmetric solutions and then fall into local optimal solutions. By adopting different initialization angles, it is ensured that neurons have high exploration and diversity during the training process. For example, in device data, if neurons have the same initialization state for responses to features such as temperature and pressure, it may lead to a single process for extracting fault features of different devices. By initializing different angles, it is equivalent to giving each neuron a "unique starting point", so that it can better adapt to complex dynamic data changes.

[0053] The second initialization angle parameter is used to control the oscillation frequency of neurons and adjust the response speed of neurons to input signals. By adjusting the second initialization angle parameter, the response range of each neuron can be controlled, making it have a stronger response under specific data patterns. For example, in device monitoring data, some sensors (such as temperature sensors) may have relatively stable responses, while other sensors (such as vibration sensors) may show more variations. Adopting the second initialization angle parameter can ensure that the responses of neurons are adaptable to different data features (such as temperature fluctuations or vibration frequency changes).

[0054] The fourth initialization angle parameter is used to control the oscillation amplitude of the neuron, affect the activation intensity of the neuron output, and ensure that the neuron can adapt to the variation amplitudes of different features in the device data. In the device data, the variation amplitudes of features (such as vibration frequency, temperature, etc.) may vary greatly. By initializing the fourth initialization angle parameter, the neuron can make adaptive adjustments according to the different amplitudes of these features, ensuring that the network can handle data variations of different scales. For example, if the temperature variation range of the device is small while the vibration frequency variation range is large, the initialization value of the fourth initialization angle parameter can adjust the response intensity of the neuron when processing different features, thereby improving the accuracy of device fault diagnosis.

[0055] 2) Based on the oscillation phase and the preset oscillation control parameters, set the initial oscillation frequency and amplitude for the neurons of the neural network.

[0056] A common problem in neural network training is that the network is prone to falling into local optimal solutions. Especially in complex device fault diagnosis tasks, the non-linear characteristics of the data make it easy for the model to find some local but not truly globally optimal solutions. To solve this problem, the present invention adopts the chaos effect. During the training process of the neuron, the chaos effect, through non-linear oscillation, enables each neuron to explore at different speeds and in different ways in each iteration cycle. The change in the oscillation frequency makes the state of the neuron not repeat in each iteration but maintain a constantly changing dynamic state, thereby increasing the breadth and depth of global search and being able to better avoid local optimal solutions. Specifically, in combination with the above oscillation phase, set the initial oscillation frequency and amplitude for each neuron, which are used to generate the chaos effect and bring dynamic search ability during the subsequent training process. The calculation methods of the initial oscillation frequency and amplitude are expressed as:

[0057]

[0058]

[0059] In the formula, is the initial oscillation frequency of the neuron, is the initial oscillation amplitude of the neuron, is the first oscillation control parameter, is the second oscillation control parameter. Preferably, = 2, = 0.1.

[0060] 3) Input the preset training sample set into the neural network to adaptively adjust the initial oscillation frequency and amplitude; and, iterate the preset oscillation phase and angle parameters based on the chaos perturbation term corresponding to the oscillation phase.

[0061] Among them, in order to enhance the exploration ability during the optimization process of neural network parameters, the embodiments of the present invention also adaptively adjust the frequency and amplitude of chaotic oscillation to cope with the situation where abnormal values appear in the operating state data of the device (such as vibration frequency or current). After the oscillation amplitude and frequency are adaptively adjusted, the search and exploration ability of the neural network can be increased. Specifically, the calculation method for adaptive adjustment of oscillation parameters is expressed as:

[0062]

[0063]

[0064] In the formula, is the oscillation amplitude of the neuron at the th iteration, is the oscillation amplitude of the neuron at the (t + 1)th iteration, is the oscillation frequency of the neuron at the tth iteration, is the oscillation frequency of the neuron at the (t + 1)th iteration; is the second initialization angle parameter of the neuron at the tth iteration, which is updated according to the gradient descent method during the training process; is the second oscillation control parameter of the neuron at the tth iteration, which is updated according to the gradient descent method during the training process. is the loss function of the neural network at the th iteration.

[0065] Among them, in the oscillation frequency update formula, the form of an exponential function is adopted, which can suppress the rapid growth of the frequency. The reciprocal of the exponential function is a rapidly decreasing function, which is used to suppress the oscillation frequency according to the loss function. When the loss function is small, the value will be large, so that the amplitude of frequency update is small, which helps to prevent the oscillation frequency of the neuron from fluctuating too violently. For example, in the task of equipment fault diagnosis, if the temperature of the equipment changes greatly (resulting in high loss), the frequency update will be faster so that the network can quickly adjust to adapt to the new data; while when the equipment data changes relatively smoothly (low loss), the speed of frequency update will slow down, thus avoiding over-adjustment and maintaining the stability of the network.

[0066] The sine function is adopted in the oscillation amplitude update formula. By introducing periodic changes, the oscillation amplitude changes periodically with the loss value during the training process. The oscillation amplitude may increase (be positive) in some stages and decrease (be negative) in other stages. This periodic adjustment helps the neural network dynamically adjust the oscillation amplitude at different training stages. For example, in device monitoring data, there may be periodic fluctuations during device operation, such as periodic changes in vibration frequency or temperature (e.g., the working cycle of mechanical equipment). The sine function enables the neural network to make adaptive adjustments in these periodic changes, allowing the neural network to better capture these periodic features and cope with occasional changes at the same time.

[0067] In summary, by using the reciprocal of the monotonically decreasing exponential function to slow down the growth rate of the frequency, as the loss decreases, the frequency update becomes more stable, which is suitable for gradually stabilizing the controlled frequency. In high-loss and unstable data patterns, it can effectively prevent over-adjustment. And the sine function is used to introduce periodic fluctuations, making the change of the amplitude periodic and fluctuating, which helps to dynamically adjust the response of neurons to different data patterns and is suitable for capturing periodic or fluctuating features existing in device data. The combined use of the two provides the flexibility of the neural network in a dynamic data environment, enabling it to extract complex features from device data more effectively and make predictions.

[0068] Among them, the loss function can be constructed based on the difference between the label obtained by performing the preset Softmax calculation on the current neural network output and the true label, which is used to measure the fitting accuracy of the neural network and update the parameters. The calculation method of the loss function is expressed as:

[0069]

[0070] In the formula, is the loss function at the t-th iteration of the neural network, is the number of samples input to the neural network in the current batch; is the true label of the i-th sample. For example, the label categories include "normal operation" and "fault status", which are represented by 0 and 1 respectively; is the label obtained by performing the preset Softmax calculation on the output of the neural network for the i-th sample, is the first oscillation control parameter of the neuron at the t-th iteration.

[0071] Furthermore, in the embodiments of the present invention, the oscillation state of neurons is changed through non-linear perturbations, avoiding being trapped in local optimal solutions during the training process and enhancing the dynamic search ability of the neural network. In the embodiments of the present invention, based on the update rule of neurons in each iteration, the oscillation phase and angle parameters are dynamically iterated, thereby dynamically iterating the oscillation state of neurons to generate non-linear perturbations in the time series. The calculation method of the oscillation phase iteration of each neuron is expressed as:

[0072]

[0073]

[0074] In the formula, is the oscillation phase of the neuron at the t-th iteration, is the first initialization angle parameter of the neuron at the t-th iteration, is the first initialization angle parameter of the neuron at the (t - 1)-th iteration, is the oscillation phase of the neuron at the (t - 1)-th iteration, is the oscillation frequency of the neuron at the t-th iteration, is the preset time step, is the chaotic perturbation term of the neuron at the t-th iteration, which is used to generate non-linear random effects in each iteration. Preferably, Δt = 0.01.

[0075] Among them, in the embodiments of the present invention, the chaotic perturbation term is calculated by non-linearly mapping the oscillation phase, the first oscillation control parameter, and the second initialization angle parameter. The first oscillation control parameter is used to control the initial growth rate of the neuron oscillation frequency, which determines the speed at which the neuron responds to input data during the training process. It affects how the neuron adjusts its frequency under different input features and data changes, and thus affects the dynamic behavior of the neuron. The second initialization angle parameter is used to control the initial phase angle of the oscillation, ensuring that the oscillation of the neuron has a certain initial state at the beginning of training. By adjusting these two parameters in the embodiments of the present invention, the neural network can respond more flexibly to different data patterns, avoid being trapped in local optimal solutions, and enhance its adaptability and search ability in a dynamic environment. Specifically, the calculation method of the chaotic perturbation term is expressed as:

[0076]

[0077] In the formula, is the perturbation coefficient, is the first oscillation control parameter of the neuron at the (t - 1)-th iteration, is the oscillation phase of the neuron at the (t - 1)-th iteration, is the second initialization angle parameter of the neuron at the (t - 1)-th iteration. Preferably, = 0.05.

[0078] Combined with the above update process, the present invention also updates the neural network weight parameters, and the update method is expressed as:

[0079]

[0080] In the formula, is the weight of the neural network at the (t + 1)-th iteration, is the weight of the neural network at the -th iteration, is the adjustment amount of the weight of the neural network based on error backpropagation at the -th iteration, is the chaotic perturbation term of the neuron at the -th iteration.

[0081] Furthermore, the calculation method of the adjustment amount of the neural network weight based on error backpropagation is expressed as:

[0082]

[0083] In the formula, is the learning rate of the neural network at the t-th iteration, is the partial derivative of the loss of the neural network with respect to the weight at the t-th iteration. Preferably, = 0.01.

[0084] Furthermore, the bias of the neural network is updated in the same way, expressed as:

[0085]

[0086] In the formula, is the bias of the neural network at the (t + 1)-th iteration, is the bias of the neural network at the t-th iteration, is the adjustment amount of the bias of the neural network based on error backpropagation at the t-th iteration.

[0087] Furthermore, the calculation method of the adjustment amount of the neural network bias based on error backpropagation is expressed as:

[0088]

[0089] In the formula, is the partial derivative of the loss of the neural network with respect to the bias at the t-th iteration.

[0090] 4) Until the neural network meets the preset training requirements, the final oscillation parameters of the neurons are obtained.

[0091] In summary, during the training process, by introducing non - linear perturbations, as the oscillation phase and frequency change, the state of the neuron changes continuously over time, avoiding fixed patterns, enhancing the network's exploration ability, and responding to fluctuations in device data, such as data on temperature changes, abnormal vibration frequencies, etc. The oscillatory state enables the neuron to adapt to the dynamic changes in the input data. In addition, the oscillatory state also characterizes the interaction between neurons. Especially when there are internal correlations between neurons, the oscillatory state can effectively capture this relationship. For example, in the data of multiple sensors of a device, the oscillatory state can help the network identify the dynamic coupling relationships between features such as temperature and vibration frequency, pressure, etc.

[0092] Step S206: Determine the self - organizing feedback amount corresponding to the neural network according to the neuron output of each neuron.

[0093] Step S208: Adjust the neuron output according to the self - organizing feedback amount to determine the target features of the operating state data.

[0094] In summary, the complex and unpredictable dynamic behavior (i.e., the oscillatory state) generated by the oscillation and chaotic effects of neurons during the training process in the embodiments of the present invention reflects the response intensity and change trend of the neurons to the input data. By adjusting the oscillation frequency and amplitude, the output of each neuron is not only related to the linear weighted result of the input device data features, but also combined with the oscillatory state of the current neuron to form a dynamic correction of the output. When the features (such as temperature, vibration, etc.) in the device data fluctuate, the oscillatory state can automatically adjust the activation output of the neuron, making it more sensitive to these fluctuations, so as to more accurately capture the abnormal state of the device. Based on this, when extracting features from device data, the real - time changes of the features (such as environmental temperature fluctuations or abnormal vibration frequencies) are dynamically coupled with the chaotic oscillatory state, facilitating the capture of the dynamic changes and non - linear features of the data, and improving the search ability of the neural network. Correspondingly, the calculation method of the activation function of the output of each neuron is expressed as:

[0095]

[0096] In the formula, is the activation output of the neuron in the neural network, is the weight of the neuron in the neural network at the t - th iteration, is the bias of the neuron in the neural network at the t - th iteration; The input device data features received by neurons in the neural network. For example, the feature attributes include: temperature, specifically referring to the ambient temperature during device operation, in degrees Celsius; vibration frequency, specifically referring to the vibration frequency during device operation, in Hertz; pressure, specifically referring to the pressure value during device operation, in Pascals; current, specifically referring to the amount of current used by the device, in Amperes; voltage, specifically referring to the voltage value used by the device, in Volts; operating duration, specifically referring to the operating time of the device from startup to the data acquisition moment, in hours; fault code, specifically referring to the error code recorded when the device fails; maintenance record, specifically referring to the time of the last maintenance of the device; production batch, specifically referring to the production batch number of the device; model, specifically referring to the model identifier of the device. Each feature attribute may exhibit different change patterns during the operation of the device, and the embodiments of the present invention can help the neural network better capture these changes based on the current oscillation state by adjusting the oscillation frequency and amplitude of the neurons. is the oscillation amplitude of the neuron at the t-th iteration, is the current iteration number, is a factor for adjusting the oscillation effect. Preferably, = 0.3. Based on this, when the features in the device data change (such as temperature fluctuations or abnormal vibration frequencies), the oscillation state can dynamically couple with the input features, automatically adjust the output of the neurons, enabling the neural network to adapt to changes in the device state, correct errors in a timely manner during the training process, dynamically correct the output, and improve the accuracy and robustness of the model.

[0097] In addition, the output of each neuron reflects their state changes, and the differences in these output states determine the feedback strength between neurons. The output of each neuron in the embodiments of the present invention is not only related to its own input signal but also related to the states of adjacent neurons. Through self-organizing feedback for collaborative adjustment of the neural network, each neuron in the neural network adjusts its output and input weights through internal local feedback and the cooperation of adjacent neurons, enabling the network to adaptively optimize its structure and tend towards an optimal configuration. Based on this, when the neural network extracts features from device data, neurons with similar device data features (such as temperature and vibration frequency may have an internal association) can use the feedback weight coefficient as a factor for adjusting the feedback strength, control the strength and importance of the feedback, affect the degree of cooperation between neurons, and achieve collaborative optimization, which can better capture the cross information of device data. The calculation method of the self-organizing feedback amount is expressed as:

[0098]

[0099] In the formula, is the feedback amount between the current neuron and the q-th neuron, is the feedback weight coefficient between the current neuron and the q-th neuron, is the output of the current neuron, is the output of the q-th neuron. Preferably, = 0.2. Further, the weights of the neural network can be adjusted based on the self-organizing feedback amount to adjust the neuron output. When specifically implemented, the update method of the neural network weight parameters implemented in the present invention is expressed as:

[0100]

[0101] is the feedback amount between the current neuron and the q-th neuron, represents the summation of the feedback amounts to all adjacent neurons of the current neuron, represents all adjacent neurons of the current neuron.

[0102] In summary, the embodiments of the present invention can introduce dynamic and irregular response patterns for the neurons of the neural network through a chaotic oscillator to prevent the network from falling into a local optimal solution during the training process. And at the beginning of network training, chaotic random initialization is performed, and unpredictability is introduced by randomizing the initial parameters (such as weights and biases) of the neurons, thereby expanding the search space and avoiding symmetric solutions. Further, during the training process, the dynamic behavior of the neurons is controlled by the set oscillation parameters (such as oscillation frequency and amplitude), so that they show non-linear changes in the time series, thereby enhancing the adaptability and search ability of the neural network to data. The embodiments of the present invention are based on the combined action of a chaotic oscillator, random initialization, and oscillation frequency and amplitude setting to ensure that the neurons can capture complex patterns in the data through complex dynamic adjustments during the training process.

[0103] For the neural network proposed in this application, by analyzing the influence of different parameter initialization methods on the model training process, comparing the convergence trends of chaotic random initialization with traditional random initialization and standard parameter initialization methods, the advantages of this technology in search space construction are verified. Referring to Figure 3 , a schematic diagram showing the influence of different initialization methods on the training loss is shown. Based on Figure 3 The experimental results shown indicate that the model processed by chaotic randomization can quickly enter the optimization track at the beginning of training, and the convergence process is more stable, indicating that the chaotic perturbation breaks the parameter symmetry, provides a more reasonable initial search direction for the neural network, effectively avoids the local optimal stagnation problem caused by unreasonable initial parameter distribution in the traditional method, and significantly shortens the training period required for the model to reach the best performance.

[0104] Further, analyze the stability performance in a noise interference environment. By injecting Gaussian noise with different intensities into the test data, compare the anti-interference capabilities of the present technology with methods such as standard neural networks, dropout regularization, and weight decay regularization. Figure 4 shows a schematic diagram of the performance comparison of different algorithms under noisy data. Refer to Figure 4 The experimental results shown indicate that as the noise intensity increases, the performance degradation of the present technology is significantly less than that of other methods. Its dynamic oscillation mechanism enables the network to spontaneously filter abnormal signals through non-linear perturbations, adaptively adjust the feature extraction strategy, maintain a stable decision boundary in a complex noise environment, and exhibit stronger environmental adaptability.

[0105] Step S210: Input the target feature into a preset fault diagnosis model to determine the fault classification result corresponding to the target feature.

[0106] Input the device data after feature extraction into a preset fault diagnosis model for classification, such as support vector machine, extreme learning machine, random forest, decision tree, Softmax function, etc., to obtain the categories of device fault diagnosis, such as including "normal operation" and "fault state", a total of 2 categories.

[0107] Step S212: Determine the fault state of the device to be tested based on the fault classification result.

[0108] In summary, another device fault diagnosis method based on sensor data provided by the embodiments of the present invention searches for operating state data based on the preset oscillation frequency and amplitude of neurons in a neural network, and uses the coupling of the oscillation state of neurons and the dynamic change of features for feature extraction, which can improve the ability to capture non-linear features of device data and improve the feature extraction effect in a complex device operating environment. Moreover, by self-organizing and feedback-adjusting the neuron output, and using the interaction between neurons to more accurately capture the internal correlation between device data features, the effectiveness of feature extraction is improved. In addition, using the chaotic randomization method to initialize the weights of the neural network and using the oscillation mechanism to set the oscillation frequency and amplitude of neurons can avoid falling into local optima, enhance the global search ability of the neural network, and effectively improve the accuracy and efficiency of device state monitoring and fault diagnosis.

[0109] Further, on the basis of the above embodiments, the embodiments of the present invention also provide another device fault diagnosis method based on sensor data. The embodiments of the present invention mainly describe the training sample set used. Figure 5 shows a flowchart of a method for constructing a training sample set provided by the embodiments of the present invention. Refer to Figure 5 The method includes the following steps:

[0110] Step S10: Obtain the pre-collected device operation monitoring data.

[0111] Step S12: According to the device operation status indicated by the device operation monitoring data, perform data annotation on the device operation monitoring data to construct an initial sample set.

[0112] Among them, the device operation monitoring data can refer to the operation status data in the above embodiments, which will not be elaborated here. Further, the collected device data is annotated to construct an initial sample set. In one embodiment, the annotation method of the present invention is manual annotation, and the annotation categories include "normal operation" and "fault status", a total of 2 categories.

[0113] It can be understood that the acquisition, annotation, and preprocessing of device data are time-consuming and laborious, and insufficient training samples are likely to lead to poor model generalization ability and affect the accuracy of the model. In the process of expanding device data in the prior art, a fixed selection method is usually used to determine the samples to be expanded. However, due to the complex distribution of device data, the prior art cannot adapt to the characteristics of complex data distribution, resulting in insufficient diversity of generated samples and a decrease in the accuracy of the classification model when dealing with device anomalies.

[0114] To adapt to the data characteristics, the embodiment of the present invention uses a dynamic selection method for data expansion, dynamically adjusting the selection range of samples to be expanded according to the feature distribution of each sample point, making the sample generation more diverse and more adaptable.

[0115] Specifically, refer to the following steps S14 - S20.

[0116] Step S14: Determine the feature distribution of the initial sample set.

[0117] The overall statistical characteristics and structure of all samples in the dataset in the feature space. The embodiment of the present invention performs data expansion based on the feature distribution, which can generate diverse and adaptable samples, enhancing the model's generalization ability for device complex distribution data. In specific implementation, the embodiment of the present invention eliminates the feature scale difference through standardization to ensure that the distance calculation (such as Euclidean distance) in the feature space can accurately reflect the true form of the data distribution. Then, according to dynamic neighborhood selection, the neighborhood range is dynamically adjusted according to the density of samples in the feature space (such as sparse or aggregated) to adapt to complex data distributions. Among them, samples have relative positions and densities in the feature space. By calculating information such as the distance, density, inter-class and intra-class distributions between sample points, the feature distribution of sample points can be comprehensively understood. Further, the embodiment of the present invention calculates the feature standard deviation of the initial sample set based on the feature mean of the initial sample set, and performs feature space normalization processing on the initial sample set based on the feature standard deviation to ensure that the expanded samples generated according to the current feature distribution are close to the actual device data distribution, avoiding the decline of training effect caused by feature scale differences.

[0118] Among them, the calculation method of the characteristic standard deviation of the original equipment data is expressed as:

[0119]

[0120] In the formula, is the total number of samples, is the characteristic value of the i-th original equipment data.

[0121] Perform normalization processing on the collected original equipment data in the feature space to ensure that the generated synthetic samples are closer to the distribution of the actual equipment data and avoid affecting the training effect of the model due to inconsistent feature scales, which is expressed as:

[0122]

[0123] In the formula, is the normalized equipment data; is the original equipment data; is the characteristic mean of the original equipment data, is the characteristic standard deviation of the original equipment data.

[0124] Step S16, determine the target position to be augmented from the initial sample set based on the feature distribution.

[0125] In specific implementation, the embodiment of the present invention determines the sample augmentation weight of the minority-class samples corresponding to the initial sample set based on the feature distribution; according to the sample augmentation weight and the Euclidean distance between each sample in the initial sample set, determine the target position to be augmented corresponding to each sample. The feature distribution contains the distribution and quantity of samples of different classes. The embodiment of the present invention mainly performs data augmentation on the minority-class samples. Minority-class samples usually cannot cover the complete feature distribution space. For example, for a characteristic value of equipment failure data (current value, the failure interval is [10A, 50A]), if there is less failure data, such as only 3 pieces of failure data, and the current values of these three pieces of failure data are set to 12A, 20A, and 25A respectively, it is obvious that the failure interval ([10A, 50A]) cannot be divided according to these three pieces of failure data. If it is augmented to 1000 pieces, the current values corresponding to these 1000 pieces of data may be spread throughout the interval [10A, 50A], then the corresponding failure interval can be well divided.

[0126] In specific implementation, the embodiment of the present invention calculates the Euclidean distance between each minority-class sample and all other samples using a dynamic distance threshold (instead of a fixed number of K nearest neighbors), and then, according to the preset dynamic distance threshold, filters out the set of neighboring samples that meet the conditions. In addition, it also preferentially selects high-weight samples for interpolation in combination with the sample weights calculated based on the reliability score.

[0127] According to the distribution of various categories in the original device data, calculate the weights of the minority class samples. Samples with higher weights will be given priority in subsequent sample generation. The weights in the embodiments of the present invention are set based on the reliability scores of the samples to emphasize the contribution of high-quality device data samples. The calculation method is expressed as:

[0128]

[0129] In the formula, is the weight of the i-th sample; is the reliability score of the i-th sample, which is obtained by calculating the class probability of the current sample through a preset Softmax classification function; is the weight adjustment threshold; is the normalization parameter. Preferably, is set to 0.2.

[0130] For each minority class sample, dynamically select the samples to be augmented according to the distribution of its feature space. The present invention adopts a dynamic proximity selection mechanism to dynamically adjust the selection range of the target augmentation position of each sample according to the feature distribution of the samples, which can better reflect the diversity of device data. In view of the complex distribution of device data, the representativeness and adaptability of the generated samples are improved. In the embodiments of the present invention, through an iterative process, the selection of the target augmentation position of each sample may be updated in each iteration to adapt to the diversity of device data. The method for dynamically selecting the target augmentation position of each sample is expressed as:

[0131]

[0132] In the formula, represents the set of target augmentation positions of the i-th sample; is the Euclidean distance calculation function; is the distance threshold; is the th normalized device data; is the th normalized device data; represents that for the th object that meets the conditions, and , is the set condition. Preferably, is set to 0.5.

[0133] Step S18, perform data augmentation on the target augmentation position through a preset sample augmentation algorithm to obtain augmented samples.

[0134] Step S20, combine the augmented samples and the initial sample set to construct a training sample set.

[0135] In one embodiment, data interpolation can be performed on the position to be augmented based on the sample augmentation weight to generate an augmented sample corresponding to the current sample. For example, the SMOTE algorithm is combined with a dynamic neighborhood selection mechanism to dynamically adjust the selection range of neighboring samples according to the feature distribution of each sample point, solving the problem that the selection of neighboring samples in the traditional SMOTE algorithm is static and the sample generation lacks diversity and adaptability.

[0136] According to the SMOTE algorithm, interpolation is performed between the selected minority-class samples and their neighboring samples to generate new synthetic samples. The way of generating new samples is expressed as:

[0137]

[0138] In the formula, is the newly generated device data sample; is the interpolation coefficient.

[0139] Among them, the interpolation coefficient can be set in a randomly generated manner to ensure the diversity of the generated samples, and at the same time avoid the generated device data samples being concentrated in certain specific areas, improving the diversity and robustness of the data, which is expressed as:

[0140]

[0141] In the formula, means following a specific distribution, means a uniform distribution on the interval [0, 1].

[0142] Furthermore, the newly generated samples are integrated with the original device data set to form an augmented device data set. In one embodiment, assume that the original collected device data samples are 800, and the device data augmentation model augments and generates 200 samples. Then the augmented device data set contains 1000 samples.

[0143] Another device fault diagnosis method based on sensor data provided by the embodiments of the present invention can dynamically adjust the selection range of the target position to be augmented according to the feature distribution of the device data through a dynamic neighborhood selection mechanism, so as to generate samples with diversity and adaptability, and enhance the generalization ability of the model to the complex distribution data of the device. Among them, combined with feature normalization processing, it can ensure that the generated samples are close to the actual device data distribution and avoid the decline of the training effect caused by the feature scale difference. Furthermore, weights are also assigned based on the reliability scores of the minority-class samples to preferentially generate high-quality samples and improve the credibility of the generated data.

[0144] Furthermore, on the basis of the above embodiments, the embodiments of the present invention also provide a device fault diagnosis device based on sensor data, Figure 6The schematic structural diagram of a device fault diagnosis device based on sensor data provided by an embodiment of the present invention is shown. Referring to Figure 6 , the device includes: a data acquisition module 100 for acquiring the operation status data of a device to be measured; a data processing module 200 for extracting chaotic randomization features from the operation status data by using a pre-trained neural network to determine the target features of the operation status data; an execution module 300 for inputting the target features into a preset fault diagnosis model to determine the fault classification result corresponding to the target features; and an output module 400 for determining the fault status of the device to be measured based on the fault classification result.

[0145] For a device fault diagnosis device based on sensor data provided by an embodiment of the present invention, its implementation principle and the technical effects produced are the same as those of the foregoing method embodiment. For the sake of brief description, for the parts not mentioned in the device embodiment, reference may be made to the corresponding content in the foregoing method embodiment.

[0146] The above data processing module 200 is further configured to input the operation status data into a preset neural network, perform chaotic random search on the operation status data based on the oscillation state of the neurons of the neural network, and determine the neuron output corresponding to each neuron of the neural network; determine the self-organization feedback amount corresponding to the neural network according to the neuron output of each neuron; adjust the neuron output according to the self-organization feedback amount to determine the target features of the operation status data.

[0147] The above data processing module 200 is further configured to obtain the oscillation parameters of the neurons of the neural network and perform chaotic random search on the operation status data based on the oscillation parameters to determine the neuron output corresponding to the neural network.

[0148] The above data processing module 200 is further configured to perform chaotic random initialization on the initial weights and biases of the neural network by using a preset chaotic function; wherein the chaotic function includes preset oscillation phase and angle parameters; set the initial oscillation frequency and amplitude for the neurons of the neural network based on the oscillation phase and a preset oscillation control parameter; input a preset training sample set into the neural network to adaptively adjust the initial oscillation frequency and amplitude; and perform iteration on the preset oscillation phase and angle parameters based on the chaotic perturbation term corresponding to the oscillation phase; until the neural network meets the preset training requirements to obtain the final oscillation parameters of the neurons.

[0149] The above data processing module 200 is further configured to update the weight parameters of the neural network based on the self-organization feedback amount to adjust the neuron output.

[0150] The device further includes a construction module, configured to obtain pre-collected device operation monitoring data; perform data annotation on the device operation monitoring data according to the device operation state indicated by the device operation monitoring data, and construct an initial sample set; determine the feature distribution of the initial sample set; based on the feature distribution, determine the target positions to be augmented in the initial sample set; perform data augmentation on the target positions to be augmented through a preset sample augmentation algorithm to obtain augmented samples; and merge the augmented samples and the initial sample set to construct a training sample set.

[0151] The above-mentioned construction module is further configured to, based on the feature distribution, determine the sample augmentation weights of the minority-class samples corresponding to the initial sample set; and determine the target positions to be augmented corresponding to each sample according to the sample augmentation weights and the Euclidean distances between each sample of the initial sample set.

[0152] The step of the above-mentioned construction module for performing data augmentation on the target positions to be augmented through a preset sample augmentation algorithm to obtain augmented samples includes: performing data interpolation on the positions to be augmented based on the sample augmentation weights to generate augmented samples corresponding to the current samples.

[0153] The above-mentioned construction module is further configured to determine the feature mean of the initial sample set, calculate the feature standard deviation of the initial sample set based on the feature mean; and perform feature space normalization processing on the initial sample set based on the feature standard deviation to determine the feature distribution of the initial sample set.

[0154] An embodiment of the present invention further provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the method shown in any of the above Figures 1 to 5 are implemented. An embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is run by a processor, the steps of the method shown in the above Figures 1 to 5 are executed. An embodiment of the present invention further provides a structural schematic diagram of an electronic device, as shown in Figure 7 which is the structural schematic diagram of the electronic device. Among them, the electronic device includes a processor 61 and a memory 60. The memory 60 stores computer-executable instructions that can be executed by the processor 61. The processor 61 executes the computer-executable instructions to implement the method shown in the above Figures 1 to 5 . In the Figure 7 shown embodiment, the electronic device further includes a bus 62 and a communication interface 63. Among them, the processor 61, the communication interface 63, and the memory 60 are connected through the bus 62.

[0155] Among them, the memory 60 may include a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk memory. The communication connection between this system network element and at least one other network element is realized through at least one communication interface 63 (which can be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. can be used. The bus 62 can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, an EISA (Extended Industry Standard Architecture) bus, etc., and can also be an AMBA (Advanced Microcontroller Bus Architecture) bus. Among them, AMBA defines three types of buses, including the APB (Advanced Peripheral Bus) bus, the AHB (Advanced High-performance Bus) bus, and the AXI (Advanced eXtensible Interface) bus. The bus 62 can be divided into an address bus, a data bus, a control bus, etc. For the sake of convenience of representation, Figure 7 only a two-way arrow is used in Figure 7 , but it does not mean that there is only one bus or one type of bus.

[0156] The processor 61 may be an integrated circuit chip with the ability to process signals. In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware in the processor 61 or the instructions in the form of software. The above-mentioned processor 61 may be a general-purpose processor, including a central processing unit (CPU for short), a network processor (NP for short), etc.; it may also be a digital signal processor (DSP for short), an application specific integrated circuit (ASIC for short), a field-programmable gate array (FPGA for short), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as being executed and completed by the hardware decoding processor, or executed and completed by a combination of the hardware and software modules in the decoding processor. The software module may be located in a mature storage medium in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor 61 reads the information in the memory and combines its hardware to complete the foregoing Figures 1 to 5 any of the shown methods.

[0157] A computer program product for a device fault diagnosis method and apparatus based on sensor data provided by an embodiment of the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the method described in the foregoing method embodiments. For specific implementation, reference can be made to the method embodiments and will not be elaborated herein. Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working process of the above-described system can refer to the corresponding process in the foregoing method embodiments and will not be elaborated herein. Additionally, in the description of the embodiments of the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", and "coupled" shall be construed in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the internal communication of two components. For those skilled in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or a part of this technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs that can store program code.

[0158] In the description of the present invention, it should be noted that the orientation or positional relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation to the present invention. In addition, the terms "first", "second", "third" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance. Finally, it should be noted that the above embodiments are only specific embodiments of the present invention, which are used to illustrate the technical solutions of the present invention rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art within the technical scope disclosed by the present invention can still modify the technical solutions recorded in the foregoing embodiments or can easily think of changes, or make equivalent replacements for some of the technical features; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A device fault diagnosis method based on sensor data, characterized in that The method includes: Obtaining the operating status data of the device to be tested; Using a pre-trained neural network to extract chaotic randomization features from the operating status data to determine the target features of the operating status data; wherein, the neural network is constructed based on self-organizing feedback; Inputting the target features into a preset fault diagnosis model to determine the fault classification result corresponding to the target features; Determining the fault status of the device to be tested based on the fault classification result; Among them, the step of using a pre-trained neural network to extract chaotic randomization features from the operating status data to determine the target features of the operating status data includes: Inputting the operating status data into a preset neural network, and performing chaotic randomization search on the operating status data based on the oscillation state of the neurons of the neural network to determine the neuron output corresponding to each neuron of the neural network; Determining the self-organizing feedback amount corresponding to the neural network according to the neuron output of each neuron; Adjusting the neuron output according to the self-organizing feedback amount to determine the target features of the operating status data; Among them, the step of inputting the operating status data into a preset neural network, and performing chaotic randomization search on the operating status data based on the oscillation state of the neurons of the neural network to determine the neuron output corresponding to each neuron of the neural network includes: Obtaining the oscillation parameters of the neurons of the neural network, and performing chaotic randomization search on the operating status data based on the neuron oscillation state corresponding to the oscillation parameters to determine the neuron output corresponding to the neural network; The calculation method of the oscillation parameters of the neurons includes: Using a preset chaotic function to perform chaotic random initialization on the initial weights and biases of the neural network; wherein, the chaotic function includes preset oscillation phase and angle parameters; Based on the oscillation phase and a preset oscillation control parameter, setting the initial oscillation frequency and amplitude for the neurons of the neural network; Inputting a preset training sample set into the neural network to adaptively adjust the initial oscillation frequency and amplitude; and iterating on the preset oscillation phase and angle parameters based on the chaotic perturbation term corresponding to the oscillation phase; Until the neural network meets the preset training requirements, obtaining the final oscillation parameters of the neurons.

2. The method according to claim 1, wherein The step of adjusting the neuron output according to the self-organizing feedback amount to determine the target features of the operating status data includes: Updating the weight parameters of the neural network based on the self-organizing feedback amount to adjust the neuron output.

3. The method according to claim 1, characterized in that, The construction method of the preset training sample set includes: Obtaining the pre-collected device operation monitoring data; According to the device operation status indicated by the device operation monitoring data, performing data annotation on the device operation monitoring data to construct an initial sample set; Determining the feature distribution of the initial sample set; Based on the feature distribution, determining the target positions to be augmented from the initial sample set; Data augmentation is performed on the target position to be augmented through a preset sample augmentation algorithm to obtain augmented samples; The augmented samples and the initial sample set are combined to construct a training sample set.

4. The method according to claim 3, wherein The step of determining the target augmentation position from the initial sample set based on the feature distribution includes: Based on the feature distribution, determining the sample augmentation weight of the minority class samples corresponding to the initial sample set; According to the sample augmentation weight and the Euclidean distance between each sample in the initial sample set, determining the target position to be augmented corresponding to each sample.

5. The method according to claim 4, characterized in that The step of performing data augmentation on the target position to be augmented through a preset sample augmentation algorithm to obtain augmented samples includes: Performing data interpolation on the position to be augmented based on the sample augmentation weight to generate augmented samples corresponding to the current samples.

6. The method according to claim 3, characterized in that, The step of determining the feature distribution of the initial sample set includes: Determining the feature mean of the initial sample set, and calculating the feature standard deviation of the initial sample set based on the feature mean; Based on the feature standard deviation, performing feature space normalization processing on the initial sample set to determine the feature distribution of the initial sample set.

7. A device fault diagnosis device based on sensor data, characterized in that, The device includes: A data acquisition module for acquiring the operating state data of the device under test; A data processing module for using a pre-trained neural network to extract chaotic random features from the operating state data to determine the target features of the operating state data; wherein, the neural network is constructed based on self-organizing feedback; An execution module for inputting the target features into a preset fault diagnosis model to determine the fault classification result corresponding to the target features; An output module for determining the fault state of the device under test based on the fault classification result; Wherein, the data processing module is further configured to: input the operating state data into a preset neural network, perform chaotic random search on the operating state data based on the oscillation state of the neurons of the neural network, and determine the neuron output corresponding to each neuron of the neural network; according to the neuron output of each neuron, determining the self-organizing feedback amount corresponding to the neural network; and adjusting the neuron output according to the self-organizing feedback amount to determine the target features of the operating state data; Wherein, the data processing module is further configured to: obtain the oscillation parameters of the neurons of the neural network, perform chaotic random search on the operating state data based on the neuron oscillation state corresponding to the oscillation parameters, and determine the neuron output corresponding to the neural network; The data processing module is further configured to: perform chaotic random initialization on the initial weights and biases of the neural network by using a preset chaotic function; wherein the chaotic function includes a preset oscillation phase and an angle parameter; set initial oscillation frequencies and amplitudes for the neurons of the neural network based on the oscillation phase and a preset oscillation control parameter; input a preset training sample set into the neural network to adaptively adjust the initial oscillation frequencies and amplitudes; and perform iteration on the preset oscillation phase and angle parameter based on a chaotic perturbation term corresponding to the oscillation phase; until the neural network meets the preset training requirements, and obtain the final oscillation parameters of the neurons.

Images