A precise air conditioner multi-source fault diagnosis method based on edge number collector

Abstract

The application provides a precise air conditioner multi-source fault diagnosis method based on an edge number collector, and belongs to the technical field of precise air conditioners. High-concurrent robust data throughput is realized through a lockless ring queue and a waiting-free read-write pointer cooperative mechanism. Then, a multi-channel data buffer is abstracted into a network flow topology graph, and memory reading bandwidth is dynamically allocated through a maximum flow minimum cut model. Riemann manifold framework-based manifold mapping is performed on multi-source heterogeneous data after scheduling, multi-source concurrent fault features are projected into a high-dimensional nonlinear orthogonal subspace, precise decoupling is realized through a spatial orthogonal decoupling topology mechanism, entropy drift indexes are calculated according to working condition entropy change rates to dynamically update edge side local weights of an artificial intelligence model, and finally, multi-source fault diagnosis results are associated and stored, and fault type identification and alarm instructions are output. The technical problem of parameter drift failure of a diagnosis model caused by feature aliasing of multi-source concurrent faults under cross-domain working condition switching is solved.

Description

Technical Field

[0001] This invention belongs to the field of precision air conditioning technology, and more specifically, relates to a method for diagnosing multi-source faults in precision air conditioning based on an edge data acquisition device. Background Technology

[0002] Precision air conditioners are core equipment in data centers, precision manufacturing, and pharmaceuticals, and their operating status directly affects the stability and reliability of the protected objects. Current mainstream solutions for precision air conditioner fault diagnosis typically employ fixed threshold detection, traditional machine learning classifiers, or static deep neural networks. These rely on manually extracted time-domain or frequency-domain statistical features to identify single fault types from multiple sensor signals, such as compressor vibration, refrigerant pressure, temperature field, and current. While these solutions offer some diagnostic capability under single stable operating conditions, in real-world environments with frequent switching between full-load in summer, low-load in winter, and variable-frequency operation during transitional seasons, the statistical distribution of sensor data shifts significantly under different conditions. Features of multi-source concurrent faults, such as refrigerant leakage and compressor malfunctions, highly overlap nonlinearly in the original feature space. Fixed-weight models cannot distinguish between distributional changes introduced by operating conditions and feature changes inherent in the fault itself. This leads to a continuous decline in diagnostic accuracy after operating condition switching, and performance cannot be restored without full retraining. In other words, existing technologies suffer from the technical problem of feature aliasing in multi-source concurrent faults during cross-domain operating condition switching, causing diagnostic model parameter drift and failure. Summary of the Invention

[0003] In view of this, the present invention provides a precision air conditioner multi-source fault diagnosis method based on edge data acquisition device, which can solve the technical problem in the prior art where feature aliasing of multi-source concurrent faults under cross-domain operating condition switching causes the diagnostic model parameters to drift and fail.

[0004] This invention is implemented as follows: This invention provides a precision air conditioning multi-source fault diagnosis method based on an edge data acquisition device, comprising the following steps:

[0005] The edge data acquisition unit collects compressor vibration, refrigerant pressure, temperature field and current signals of precision air conditioners. Through the lockless ring queue and waitless read and write pointer collaborative mechanism, it completes high-concurrency and robust throughput of multi-source high-frequency streaming data and outputs multi-channel time-series data buffer.

[0006] The multi-channel time-series data buffer is abstracted into a network flow topology graph with limited capacity. The memory read bandwidth and computation priority of each sensor data stream are dynamically allocated by solving the maximum flow minimum cut model.

[0007] Manifold mapping is performed on the multi-source heterogeneous data that has been scheduled, projecting the original features onto a high-dimensional nonlinear orthogonal subspace, establishing a spatial orthogonal decoupling topology mechanism, and achieving precise decoupling and isolation of multi-source concurrent fault features;

[0008] The decoupled standardized spatiotemporal tensor is input into the dynamic reconstruction of the local activation space evolution model. The model dynamically reconstructs the hidden layer operator parameters based on the local topological distribution of the input data, and the distribution offset is detected by a bidirectional backtracking gating mechanism. The memory trace decay branch is called to perform incremental correction on the layer weight matrix and output multi-source fault diagnosis results.

[0009] The entropy drift adaptive adjustment function is calculated based on the entropy change rate under the operating conditions to obtain the entropy drift index. The edge local weights of the dynamically reconstructed local activation space evolution model are then dynamically updated based on the interval to which the entropy drift index belongs.

[0010] The multi-source fault diagnosis results are associated with and stored with the equipment operating status records, and the fault type identifier and alarm command are output through the edge data acquisition device.

[0011] The lock-free circular queue refers to replacing the traditional locked queue with a fixed-capacity circular buffer. The producer and consumer threads each hold independent write pointers and read pointers, and there is no mutual exclusion waiting between the two pointers. The queue capacity is determined by the product of the measured peak sampling rate and the inference latency, and the packet loss rate is adjusted to below an acceptable threshold through iterative stress testing.

[0012] The aforementioned wait-free read-write pointer coordination mechanism uses atomic operations to advance both the write pointer and the read pointer, allowing any thread to complete read and write operations directly at any time without waiting for other threads to release resources.

[0013] The network flow topology graph with limited capacity is a directed capacity network with each sensor channel as the source node, the input layer of the dynamic reconstruction local activation space evolution model as the sink node, and the current occupancy and remaining bandwidth of each channel's data buffer as the edge capacity. The maximum flow minimum cut model is solved by the Ford-Fulkerson algorithm, and the edge set corresponding to the minimum cut is the bottleneck edge set of the memory bandwidth at the current moment. The determination of high-frequency key fault feature data flow is based on the frequency band where the fault feature energy is concentrated.

[0014] The manifold mapping utilizes a nonlinear dimensionality reduction operator based on the Riemannian manifold framework, using the geodesic distance of the covariance matrix as a feature similarity measure. By performing singular value decomposition on the geodesic distance matrix of historical fault data, the orthogonal projection direction is determined, and multiple fault features are mapped to a high-dimensional nonlinear orthogonal subspace.

[0015] The spatial orthogonal decoupling topology mechanism establishes isolation boundaries based on the topological connectivity of the characteristic distribution of each fault type, so that the topological neighborhoods of the refrigerant leakage characteristic submanifold and the compressor abnormal characteristic submanifold in the subspace do not intersect.

[0016] The dynamic reconstruction local activation space evolution model has an input layer receiving dimension of... The standardized spacetime tensor, in which Number of sensor channels The time step is specified; after the input layer is a streaming dynamic threshold normalization layer; followed by... The hidden layer consists of stacked local activation operators; the hidden layer is equipped with bidirectional backtracking gating units; the output layer is a fully connected classification layer.

[0017] Wherein, the internal maintenance dimension of the local activation operator is The local prototype matrix is ​​calculated during forward propagation, and the cosine similarity between the input feature vector and each prototype is calculated. The similarity is used as a soft gating weight to perform weighted aggregation of the prototype matrix, so as to realize the dynamic reconstruction of the operator parameters with the local topological distribution of the input. The spatial self-organizing evolution mechanism of the prototype matrix moves the prototype towards the local density center of the input data through competitive learning during the training phase.

[0018] The bidirectional backtracking gating unit monitors the entropy value of the hidden state distribution of adjacent layers in real time. When the entropy of the hidden state exceeds the entropy surge threshold, the main forward path is cut off and the process jumps to the memory trace decay branch. The memory trace decay branch maintains the exponentially decaying historical state cache matrix and uses the historical state cache matrix to perform incremental correction on the current layer weight matrix. The correction amount is calculated by the cross product of the historical state cache matrix and the current hidden state.

[0019] The update method of the exponentially decaying historical state cache matrix in the memory trace decay branch refers to the update of the cache matrix after each inference step using a decay coefficient. Multiply by the current cache matrix, and add to Multiply by the current hidden state matrix to make the contribution of the recent state to the incremental correction greater than that of the long-term state.

[0020] The network loss function of the dynamically reconstructed local activation space evolution model consists of two parts: cross-entropy classification loss and physical constraint penalty term. The physical constraint penalty term transforms the energy conservation equation and mass conservation equation of the precision air conditioning system into a residual penalty. Adjacent hidden layers maintain consistent feature tensor dimensions through an adaptive step-size tensor splicing operator. The splicing operator is based on the output tensor of the current layer. The norm adaptively adjusts the step size.

[0021] In this process, the memory allocation of the dynamically reconstructed local activation space evolution model uses the bottleneck edge set capacity value of the maximum flow minimum cut model as the memory allocation ratio coefficient of each hidden layer; the inter-layer neuron connection density is adaptively pruned based on the gate weight variance output by the local activation operator of the previous layer; and the batch size is dynamically adjusted based on the real-time occupancy rate of the lockless circular queue.

[0022] The entropy drift adaptive adjustment function uses the weighted average of the absolute values ​​of the differences between the Shannon entropy of each operating parameter and the Shannon entropy of the corresponding parameter in the previous time window as the entropy drift index. ;when When the weights are below the first entropy drift threshold, the local weights on the edge side remain unchanged; when... When the weights are between the first entropy drift threshold and the second entropy drift threshold, perform a single-step gradient descent incremental update on the local weights on the edge side with a first-step length; when... When the value is not lower than the second entropy drift threshold, the memory trace decay branch is activated to perform multi-step incremental correction on all hidden layer weight matrices with a second step size.

[0023] The training of the dynamic reconstruction local activation spatial evolution model involves simulating refrigerant leakage, compressor bearing wear, compressor coil overheating, and concurrent combined faults on a precision air conditioning test bench. At least 72 hours of multi-source sensor data are collected under different seasonal operating conditions. A balanced sample set is formed by dividing the data into training, validation, and test sets according to their proportions. The model parameters are iteratively updated using the AdamW optimizer. After each training round, the entropy surge threshold is adjusted based on the entropy distribution of the validation set. Training stops when the validation accuracy does not improve after consecutive rounds.

[0024] Among them, time step The value range is 64 to 512; the number of local activation operator layers The value range is 4 to 12; number of prototypes The value range is 8 to 64; attenuation coefficient The value range is 0.85 to 0.99; the step size range of the adaptive step size tensor splicing operator is 1 to 4; the first step size... The range of values ​​is ~ Second step length The range of values ​​is ~ And the second step length Greater than the length of the first step The initial learning rate range for the AdamW optimizer is... ~ .

[0025] This invention projects multi-source concurrent fault features onto a high-dimensional nonlinear orthogonal subspace through manifold mapping and spatial orthogonal decoupling topology mechanisms. This ensures that the feature submanifolds of refrigerant leakage and compressor anomalies are topologically disjoint, fundamentally eliminating the interference of feature aliasing on diagnosis. Based on this, a dynamically reconstructed local activation space evolution model dynamically reconstructs the hidden layer operator parameters according to the local topological distribution of the input data. A bidirectional backtracking gating mechanism automatically switches to the memory trace decay branch and performs incremental correction on the weights when a distribution shift is detected, allowing the model to recover diagnostic performance without retraining after cross-domain operating condition switching. An entropy drift adaptive adjustment function further dynamically updates the local weights on the edge side according to the entropy change rate of the operating condition, suppressing the cumulative effect of parameter drift. In summary, this invention solves the technical problem mentioned in the background art where feature aliasing in multi-source concurrent faults leads to parameter drift failure of the diagnostic model under cross-domain operating condition switching. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention.

[0027] Figure 2 This is a comparison diagram of the topological distribution of multi-source fault characteristics in the orthogonal subspace and the original space.

[0028] Figure 3 This is a graph showing the trend of entropy drift index as the operating conditions change and the graded trigger response. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0030] like Figure 1 The diagram shown is a flowchart of a multi-source fault diagnosis method for precision air conditioning based on an edge data acquisition device provided by the present invention. This method includes the following steps:

[0031] S01, the edge data acquisition unit collects the compressor vibration, refrigerant pressure, temperature field and current signals of the precision air conditioner, and completes the high-concurrency and robust throughput of multi-source high-frequency streaming data through the lockless ring queue and the waitless read and write pointer collaborative mechanism, and outputs a multi-channel time-series data buffer.

[0032] S02. The multi-channel time-series data buffer is abstracted into a network flow topology graph with limited capacity. The memory read bandwidth and calculation priority of each sensor data flow are dynamically allocated by solving the maximum flow minimum cut model to ensure that the high-frequency key fault feature data flow enters the subsequent processing path first.

[0033] S03. Perform manifold mapping on the multi-source heterogeneous data that has been scheduled, project the original features onto a high-dimensional nonlinear orthogonal subspace, establish a spatial orthogonal decoupling topology mechanism, and achieve precise decoupling and isolation of multi-source concurrent fault features such as refrigerant leakage and compressor abnormality.

[0034] S04. Input the decoupled standardized spatiotemporal tensor into the dynamic reconstruction of the local activation space evolution model. The model dynamically reconstructs the hidden layer operator parameters based on the local topological distribution of the input data. The distribution offset is detected by a bidirectional backtracking gating mechanism. The memory trace decay branch is called to perform incremental correction on the layer weight matrix and output the multi-source fault diagnosis results.

[0035] S05. Calculate the entropy drift adaptive adjustment function based on the entropy change rate under the operating conditions to obtain the entropy drift index. Dynamically update the edge local weights of the dynamically reconstructed local activation space evolution model according to the interval to which the entropy drift index belongs, so as to suppress parameter drift and performance degradation under cross-domain operating conditions.

[0036] S06. The multi-source fault diagnosis results are associated with and stored with the equipment operation status records. The fault type identifier and alarm command are output through the edge data acquisition device to complete the closed-loop diagnosis of multi-source faults in the precision air conditioner.

[0037] The lock-free circular queue refers to replacing the traditional locked queue with a fixed-capacity circular buffer when multiple threads collect high-frequency streaming data in parallel. Producer and consumer threads each hold independent write pointers and read pointers, and there is no mutual exclusion or waiting between the two pointers, thereby completely eliminating the data packet loss and real-time inference interruption problems caused by shared memory lock contention. The queue capacity is determined by the product of the measured peak sampling rate and the inference latency. Through continuous stress testing at different sampling frequencies and statistical analysis of packet loss rates, the capacity parameters are iteratively adjusted until the packet loss rate is lower than an acceptable threshold. The acceptable threshold is determined by analyzing the cost of fault missed detection under the rated operating conditions of a precision air conditioner.

[0038] The aforementioned wait-free read-write pointer coordination mechanism means that both the write pointer and the read pointer are advanced using atomic operations. Any thread can directly complete the read and write operations at any time without waiting for other threads to release resources. When coordinated with a lock-free circular queue, it can achieve extremely low latency multi-source data throughput under the condition of limited edge data acquisition device resources.

[0039] The network flow topology graph with limited capacity is established as follows: each sensor channel of the edge data acquisition unit is used as the source node, the input layer of the dynamic reconstruction local activation space evolution model is used as the sink node, and the current occupancy and remaining bandwidth of each channel's data buffer are used as the edge capacity to form a directed capacity network. The maximum flow minimum cut model is solved on the directed capacity network using the Ford-Fulkerson algorithm. The edge set corresponding to the minimum cut is the bottleneck edge set of the memory bandwidth at the current moment. The algorithm dynamically increases the memory read priority of the high-frequency key fault feature data stream based on the bottleneck edge set, while reducing the priority of the low-frequency auxiliary data stream, thus completely eliminating the memory blocking bottleneck under multi-source high concurrency. The determination criteria for the high-frequency key fault feature data stream are as follows: in the previous experiment, the characteristic frequency bands of typical fault modes of precision air conditioners (refrigerant leakage, compressor abnormality) were analyzed. Channels with fault feature energy concentrated above 50Hz were identified as high-frequency key fault feature data streams. The 50Hz threshold was determined by cross-validation experiments on the fault detection rate at different cutoff frequencies.

[0040] The manifold mapping refers to using a nonlinear dimensionality reduction operator to map highly nonlinearly overlapping multi-fault features in the original high-dimensional sensor feature space to a higher-dimensional nonlinear orthogonal subspace, so that the feature distributions of different fault modes are orthogonal to each other in the subspace, thereby achieving precise decoupling and isolation of multi-source concurrent faults. The mapping operator is based on the Riemannian manifold framework and uses the geodesic distance of the covariance matrix as a feature similarity measure. The orthogonal projection direction is determined by performing singular value decomposition on the geodesic distance matrix of the historical fault data of precision air conditioners.

[0041] The spatial orthogonal decoupling topology mechanism refers to establishing isolation boundaries in a high-dimensional nonlinear orthogonal subspace based on the topological connectivity of the characteristic distribution of each fault type, so that the topological neighborhoods of the refrigerant leakage characteristic submanifold and the compressor abnormal characteristic submanifold in the subspace do not intersect, thereby achieving stable decoupling of concurrent faults.

[0042] The dynamic reconstruction local activation space evolution model is based on deep learning, and its specific structure is as follows: the model input layer receives data in dimensions of... The standardized spacetime tensor, in which Number of sensor channels For time step, The value range is 64–512, and the specific value is determined through experimental evaluation of the integrity of fault characteristics under different time windows; after the input layer is a streaming dynamic threshold normalization layer, which updates the normalization parameters in real time based on the mean and variance of the current batch of data to eliminate the data distribution offset introduced by seasonal operating condition switching; after the streaming dynamic threshold normalization layer is Hidden layers formed by stacking local activation operators The value ranges from 4 to 12, and the specific value is determined by Pareto analysis of diagnostic accuracy and inference latency at different levels on the validation set; each local activation operator is driven by a spatial self-organizing evolution mechanism, specifically implemented as follows: the operator internally maintains a dimension of The local prototype matrix, where For the number of prototypes, For feature dimension, The value range is 8–64. During forward propagation, the operator calculates the cosine similarity between the input feature vector and each prototype. This similarity is used as a soft-gating weight to weight and aggregate the prototype matrix, resulting in a local activation output. This achieves dynamic reconstruction of the operator parameters based on the local topological distribution of the input, rather than relying on fixed weights. A bidirectional backtracking gating unit is set in the hidden layer. This unit monitors the entropy of the hidden layer state distribution in real time. When the hidden layer state entropy exceeds the entropy surge threshold, the gating unit cuts off the main forward path and jumps to the memory trace decay branch. The memory trace decay branch maintains an exponentially decaying historical state cache matrix with a decay coefficient... The value ranges from 0.85 to 0.99, and the specific value is determined by applying different... The parameter drift of the model on a cross-seasonal working condition dataset was determined through multiple rounds of experiments. Branches incrementally correct the current layer weight matrix using a historical state cache matrix; the correction amount is calculated by the outer product of the historical state cache matrix and the current hidden layer state. The network loss function consists of two parts: cross-entropy classification loss and a physical constraint penalty term. The physical constraint penalty term transforms the energy conservation equation and mass conservation equation of the precision air conditioning system into a residual penalty, ensuring that the model follows the physical mechanism when extrapolating in the feature space. Adjacent hidden layers maintain consistent feature tensor dimensions through an adaptive step-size tensor concatenation operator, which is based on the output tensor of the current layer. The norm adaptively adjusts the step size, ranging from 1 to 4; the output layer is a fully connected classification layer, outputting the probability distribution of each fault type; model memory allocation adopts a graph theory-based hierarchical bandwidth allocation strategy, specifically: the bottleneck edge set capacity of the maximum flow minimum cut model is used as the memory allocation ratio coefficient for each hidden layer, with layers having larger bottleneck bandwidth allocated more memory to support wider local prototype matrices, realizing the linkage calculation relationship between memory resources and network flow scheduling parameters; the inter-layer neuron connection density is adaptively pruned based on the gate weight variance output of the previous layer's local activation operator, reducing the number of connections to reduce redundant computation when the variance is small, and the variance threshold is determined through quantile analysis of the gate weight variance distribution on the training set; the batch size in the data loop is dynamically adjusted based on the real-time occupancy rate of the current lock-free circular queue, increasing the batch size when the occupancy rate is high to accelerate processing. To reduce latency, the batch size is reduced when the occupancy rate is low. The steps for establishing the training dataset for the dynamically reconstructed local activation spatial evolution model specifically include: simulating refrigerant leakage, compressor bearing wear, compressor coil overheating, and concurrent combinations of the above faults on a precision air conditioning test bench; collecting at least 72 hours of multi-source sensor data under different seasonal operating conditions (summer full load, winter low load, and transitional season frequency conversion); and having professional technicians perform dual labeling of fault type and operating condition type on each data segment to form a balanced sample set containing single faults, concurrent faults, and normal operating conditions, which is divided into training set, validation set, and test set in a 7:1:2 ratio. The steps for training the dynamically reconstructed local activation spatial evolution model specifically include: iteratively updating the model parameters using the AdamW optimizer, with an initial learning rate range of [value missing]. ~ The specific values ​​are determined by the learning rate warm-up experiment on the validation set loss curve to determine the optimal convergence interval; after each training round, the hidden layer state entropy distribution is evaluated on the validation set, and the entropy surge threshold of the bidirectional backtracking gating unit is adjusted according to the entropy distribution; when the diagnostic accuracy of the validation set does not improve within 10 consecutive rounds, training is stopped and the optimal weights are saved.

[0043] The technical effects of the dynamically reconstructed local activation space evolution model on the entire scheme are as follows: the local activation operator dynamically reconstructs the operator parameters based on the local topology distribution of the input data, enabling the network to adaptively adjust its internal representation space rather than relying on fixed weights when facing cross-domain conditions, fundamentally suppressing cross-domain parameter drift; the bidirectional backtracking gating mechanism automatically switches to the memory trace decay branch when a distribution shift is detected, using historical states to incrementally correct the weights, allowing the model to quickly recover its diagnostic performance without retraining after a change in conditions; the physical constraint penalty term embeds energy conservation and mass conservation into the loss function, constraining the model to always follow the physical mechanism of a precision air conditioning system when extrapolating the feature space, significantly improving the model's generalization ability to unseen conditions; the linkage mechanism between network flow topology scheduling and memory allocation ensures efficient utilization of edge-side resources, achieving robust diagnosis of multi-source concurrent faults with minimal computational overhead.

[0044] The specific calculation method of the entropy drift adaptive adjustment function is as follows: the weighted average of the absolute values ​​of the differences between the Shannon entropy of each operating parameter (compressor speed, evaporating pressure, condensing pressure, ambient temperature) of the precision air conditioner in the current time window and the Shannon entropy of the corresponding parameter in the previous time window is used as the entropy drift index. The weights of each parameter were determined experimentally using the principal component contribution rate of historical operating condition switching data; when At this time, the local weights on the edge side of the model remain at their current values ​​and are not updated. Statistical analysis of fluctuation data under normal operating conditions on the test bench The 95th percentile is determined; when At that time, by step length Perform a single-step gradient descent incremental update on the local weights at the edge. The range of values ​​is ~ ;when At that time, the memory trace decay branch is activated, with a step size. Perform multi-step incremental corrections on all hidden layer weight matrices. The range of values ​​is ~ , By statistically analyzing the seasonal operating condition switching experiments on the test bench, we can determine the corresponding values ​​that cause a decrease in diagnostic accuracy exceeding acceptable losses. The quantiles are determined, and the acceptable loss is given by the cost analysis of missed detections in precision air conditioning systems; all of the above intervals are... type of half-open interval, and They are represented using inequality signs respectively.

[0045] The aforementioned spatial self-organizing evolution mechanism refers to the fact that the prototype matrix inside the local activation operator moves self-organizing towards the local density center of the input data through competitive learning during the training phase. During the inference phase, the prototype matrix is ​​fixed, but the soft gating weights are dynamically calculated with each input, enabling the operator to achieve a dynamic response to changes in the input distribution without changing the number of parameters.

[0046] The update method for the exponentially decaying historical state cache matrix in the memory trace decay branch is as follows: after each inference step, the cache matrix is ​​updated with a decay coefficient. Multiply by the current cache matrix, and add to Multiply by the current hidden state matrix to cause the influence of historical states to decay exponentially with the number of time steps, and the recent state contributes more to the incremental correction than the long-term state.

[0047] The streaming dynamic threshold management refers to the edge data acquisition device dynamically adjusting the anomaly detection threshold based on the statistical characteristics (mean, variance, kurtosis) of the real-time data of each channel during the multi-source high-frequency data acquisition process, so as to avoid false alarms or missed alarms under fixed thresholds due to the switching of operating conditions.

[0048] The operating condition entropy change rate refers to the rate of change of Shannon entropy of various operating parameters of the precision air conditioner per unit time. It is used to quantify the intensity of the switching of the equipment's operating state and is the core input of the entropy drift adaptive adjustment function.

[0049] The specific technical effect of the Ford-Fulkerson algorithm in this scheme is as follows: by repeatedly searching for augmenting paths and updating residual capacity on the directed capacity network, the algorithm converges to the maximum flow value within a finite number of iterations. The corresponding minimum cut edge set directly indicates the current memory bandwidth bottleneck position, making the scheduling decision globally optimal rather than a heuristic approximation. This ensures that the high-frequency critical fault feature data stream can obtain sufficient memory read bandwidth under any concurrent state, fundamentally eliminating the memory blocking bottleneck when multi-threaded concurrent acquisition is performed on the edge side.

[0050] The adaptive step-size tensor concatenation operator functions as follows: between adjacent hidden layers, based on the output tensor of the previous layer... The norm adaptively selects the concatenation step size. When the step size is large, the concatenation is skipped to compress redundant features, and when the step size is small, the concatenation is dense to retain fine-grained features, ensuring that the feature tensor has consistent dimensions and minimal information loss when it is passed across layers.

[0051] Optionally, this solution also provides a computer-based precision air conditioner multi-source fault diagnosis system. The computer is equipped with a readable storage medium, which stores program instructions. When the program instructions are run in the computer, they combine an edge data acquisition unit, a lock-free circular queue buffer, and a dynamically reconstructed local activation space evolution model to execute all the steps of the above-mentioned precision air conditioner multi-source fault diagnosis method.

[0052] The specific implementation of step S01 is as follows: The edge data acquisition unit synchronously acquires the compressor vibration acceleration, refrigerant high and low pressure sides, evaporator and condenser temperature fields, and compressor drive current of the precision air conditioner through an analog-to-digital conversion interface. The sampling rate is set according to the fault characteristic frequency band: the vibration channel is no less than 1000Hz, the pressure and temperature channel is no less than 100Hz, and the current channel is no less than 500Hz. Multi-channel acquisition threads run in parallel, each thread independently holding a write pointer. The write pointer is advanced atomically, and the sampled value is written to a fixed-capacity circular buffer without the need for a mutex lock, thus eliminating lock contention during multi-threaded writes. The consumer thread also advances the read pointer atomically, and there is no waiting relationship between it and the production thread, achieving a wait-free read-write pointer coordination mechanism. The circular queue capacity is determined by the product of the measured peak sampling rate and the subsequent inference delay. The reference value is twice the product of the peak sampling rate and the inference delay to ensure no packet loss occurs during inference jitter. The acceptable packet loss rate threshold is set through analysis of the cost of fault missed detection under the rated operating conditions of the precision air conditioner, and is typically below 0.1%. The purpose of this step is to achieve robust concurrent acquisition and low-latency buffering of multiple high-frequency sensor signals under the condition of limited computing resources of edge data acquisition devices, so as to provide a complete multi-channel time-series data buffer for subsequent scheduling and diagnosis.

[0053] The specific implementation of step S02 is as follows: Using each sensor channel as the source node, the input layer of the dynamically reconstructed local activation space evolution model as the sink node, and the current occupancy of each channel's data buffer and remaining memory bandwidth as the edge capacity, a directed capacity network is constructed, forming a capacity-constrained network flow topology. The Ford-Fulkerson algorithm is run on the directed capacity network, repeatedly searching for augmenting paths from the source node to the sink node and updating the residual capacity. The minimum cut edge set corresponding to the maximum flow value obtained after algorithm convergence is the memory bandwidth bottleneck edge set at the current moment. The determination criterion for high-frequency critical fault feature data flow is: in previous experiments, power spectrum analysis was performed on the characteristic frequency bands of typical fault modes in precision air conditioners, and channels with fault feature energy concentrated above 50Hz were identified as high-frequency critical fault feature data flows. Based on the bottleneck edge set, the scheduling algorithm increases the memory read priority of the high-frequency critical fault feature data flow while decreasing the priority of the low-frequency auxiliary data flow, ensuring that critical feature data enters the subsequent processing path first. The purpose of this step is to use the maximum flow minimum cut theory to perform globally optimal allocation of limited memory bandwidth on the edge side, eliminating memory blocking bottlenecks under multi-source high concurrency.

[0054] The specific implementation of step S03 is as follows: For the multi-source heterogeneous data that has been scheduled, a nonlinear dimensionality reduction operator based on the Riemannian manifold framework is constructed. First, the covariance matrix of each sensor channel within the time window is used as the feature representation, and the geodesic distance between the covariance matrices is used as the feature similarity measure. The full geodesic distance matrix is ​​calculated based on the historical fault dataset. Singular value decomposition is performed on the geodesic distance matrix, and the first few singular vectors are selected as orthogonal projection directions to project the multi-fault features in the original high-dimensional sensor feature space to a higher-dimensional nonlinear orthogonal subspace. In this subspace, isolation boundaries are established based on the topological connectivity of the feature distribution of each fault type, so that the topological neighborhoods of the refrigerant leakage feature submanifold and the compressor abnormality feature submanifold do not intersect, thus forming a spatial orthogonal decoupling topological mechanism to achieve accurate decoupling and isolation of multi-source concurrent fault features. The purpose of this step is to eliminate the nonlinear overlap of multi-source fault features in the low-dimensional space and provide topologically separable feature inputs for subsequent models.

[0055] The specific implementation of step S04 is as follows: the decoupled features are standardized to form a dimension of The spatiotemporal tensor is input into a dynamically reconstructed local activation space evolution model. The model first passes through a streaming dynamic threshold normalization layer, which updates the normalization parameters in real time based on the mean and variance of the current batch, eliminating data distribution bias introduced by seasonal operating condition switching. Then it proceeds to... A hidden layer consisting of stacked local activation operators, each local activation operator internally maintaining... A local prototype matrix is ​​used. During forward propagation, the cosine similarity between the input feature vector and each prototype is calculated. This similarity is used as a soft-gating weight to weight and aggregate the prototype matrix, achieving dynamic reconstruction of operator parameters according to the local topological distribution of the input. In the hidden layer, a bidirectional backtracking gating unit monitors the Shannon entropy of the hidden layer states of adjacent layers in real time. When the entropy value exceeds the entropy surge threshold, the main forward path is cut off, and the memory trace decay branch is activated. Within the branch, the decay coefficient is used to... The model maintains an exponentially decaying historical state cache matrix and calculates the incremental correction by multiplying the historical state cache matrix with the current hidden state. This incremental correction is then applied to the current layer weight matrix, enabling the model to quickly recover its diagnostic performance after a change in operating conditions. The model output layer is a fully connected classification layer, outputting the probability distribution of each fault type, i.e., the multi-source fault diagnosis results. The goal of this step is to achieve adaptive and robust diagnosis of multi-source concurrent faults without relying on full retraining.

[0056] The specific implementation of step S05 is as follows: The entropy drift index is obtained by calculating the weighted average of the absolute values ​​of the differences between the Shannon entropy of each parameter (compressor speed, evaporating pressure, condensing pressure, and ambient temperature) within the current time window and the corresponding Shannon entropy of the parameter in the previous time window. The weights of each parameter were determined experimentally using the principal component contribution rate of historical operating condition switching data. Based on the entropy drift adaptive adjustment function, when... Below the first entropy drift threshold At that time, the local weights on the edge side remain unchanged. The reference value is from the fluctuation data of the test bench under normal operating conditions. 95th percentile; when Between With the second entropy drift threshold In between, with the first step length (Reference range) ~ Perform single-step gradient descent incremental update; when Not less than At that time, the memory trace decay branch is activated, with a second step length. (Reference range) ~ Perform multi-step incremental corrections on all hidden layer weight matrices. The reference value is the value corresponding to the decrease in diagnostic accuracy exceeding acceptable loss during seasonal operating condition switching experiments. The quantiles. The purpose of this step is to trigger weight updates in stages based on the severity of the operating condition transition, thereby suppressing parameter drift and performance degradation under cross-domain operating conditions.

[0057] The specific implementation of step S06 is as follows: The multi-source fault diagnosis results output by the dynamically reconstructed local activation space evolution model, including the fault type probability distribution and judgment identifier, are structurally associated and stored with the current equipment operating status record (operating parameter snapshot, sensor raw data index) to form a traceable fault diagnosis log. Based on the fault type judgment result, the edge data acquisition unit generates the corresponding fault type identifier code and sends an alarm command to the upper-level system or field alarm device to complete the closed-loop diagnosis of multi-source faults in the precision air conditioner. The purpose of this step is to associate the diagnosis results with the operating status, support post-event traceability analysis, and achieve integration with the upper-level system through a standardized alarm output interface.

[0058] It should be noted that the key technologies of this invention include: a manifold mapping and spatial orthogonal decoupling topology mechanism based on the Riemannian manifold framework, which determines the orthogonal projection direction through the singular value decomposition of the geodesic distance matrix, making the topological neighborhoods of multi-source concurrent fault features non-intersecting in a high-dimensional nonlinear orthogonal subspace, thus physically blocking feature aliasing and providing a topologically separable feature foundation for subsequent models. The local prototype matrix of the dynamically reconstructed local activation space evolution model is dynamically reconstructed using a cosine similarity soft gating mechanism, combined with a bidirectional backtracking gating unit and a memory trace decay branch, enabling the model to automatically switch to the incremental correction path when a distribution shift is detected, restoring diagnostic performance without full retraining. The entropy drift adaptive adjustment function triggers local weight updates on the edge side in stages according to the rate of change of the Shannon entropy parameter, actively suppressing parameter drift. The three key technologies mentioned above work together: manifold mapping eliminates aliasing at the feature level, the dynamic reconstruction mechanism of the model addresses the distribution shift at the representation level, and the entropy drift adjustment function forms a closed-loop feedback at the working condition perception level. Together, they constitute a full-link anti-drift mechanism from feature input to weight update, enabling the present invention to have robust multi-source concurrent fault diagnosis capabilities under cross-domain working conditions. Such robustness cannot be achieved by any single technical component alone.

[0059] It should be noted that under the condition of long-term continuous operation of precision air conditioners spanning multiple seasons, the statistical distribution of parameters such as condensing pressure, evaporating pressure, and ambient temperature will undergo a continuous overall shift with seasonal changes. This shift is not a local disturbance but a global distribution domain migration. In this case, there is a significant statistical difference between the hidden layer activation distribution obtained by the model under full-load conditions in summer and the actual activation distribution under low-load conditions in winter. This leads to a continuous increase in hidden layer state entropy, the classification decision surface of the main forward path deviating from the true fault boundary, and the diagnostic accuracy monotonically declining over time. The reason for the above technical problem is that the weights of the neural network remain fixed after training, while the statistical distribution of sensor data changes continuously with the operating conditions. The mismatch between the fixed weights and the dynamic distribution is reflected in the continuous drift of the activation distribution in the hidden layer. The feature extraction capability of the network gradually becomes mismatched with the data manifold of the current operating conditions, eventually causing the probability output of the classification layer to become unreliable. The usual solution to the above technical problem is to periodically retrain the model with a full dataset. That is, after each change in operating conditions, sufficient labeled data is collected again and all parameters are optimized from scratch or from the pre-trained weights. However, full retraining requires a large number of labeled samples and high-intensity computing resources, while the computing power of edge data acquisition devices is extremely limited, making it impossible to complete full backpropagation online. Simultaneously, failure events in precision air conditioners are rare in actual operation, and labeled failure samples under new operating conditions are often insufficient. Full retraining is prone to overfitting to new operating conditions or forgetting old operating conditions, making it difficult to control training costs while ensuring diagnostic accuracy in engineering. This invention effectively solves this technical problem. The bidirectional backtracking gating unit continuously monitors the entropy values ​​of adjacent hidden layer state distributions. Once it detects that the hidden layer state entropy exceeds the entropy surge threshold, it immediately cuts off the main forward path relying on fixed weights and jumps to the memory trace decay branch. The memory trace decay branch performs incremental correction on the current layer weight matrix by exponentially decaying the historical state cache matrix. The correction amount only involves the outer product of the current hidden layer state and the historical state cache matrix, with a computational cost far less than full backpropagation, which can be completed in real time at the edge. (Attenuation coefficient) The design ensures that recent states contribute more to weight adjustments than long-term states, avoiding over-reliance on historical operating conditions. Simultaneously, the entropy drift adaptive adjustment function triggers weight updates in stages based on the rate of entropy change under operating conditions. When operating conditions are stable, weights are frozen to prevent unnecessary disturbances, while the adjustment step size is increased to quickly track distribution changes during drastic shifts in operating conditions. These two mechanisms work synergistically, enabling the model to achieve continuous and stable diagnosis across seasonal operating conditions through lightweight online incremental adjustments without requiring full retraining.

[0060] Specifically, the principle of this invention is:

[0061] The fundamental reason why this invention can solve the above-mentioned technical problems is that it works synergistically from three levels: feature decoupling, model adaptive reconstruction, and working condition perception, forming a complete anti-drift closed-loop mechanism.

[0062] At the feature decoupling level, the features of multi-source concurrent faults in the original sensor signals exhibit high nonlinear overlap in the low-dimensional space, which is the root cause of the confusion between different fault types in the fixed-weight model. This invention employs a manifold mapping operator based on the Riemannian manifold framework, using the geodesic distance of the covariance matrix as a feature similarity measure. By performing singular value decomposition on the geodesic distance matrix of historical fault data to determine the orthogonal projection direction, the original features are projected onto a high-dimensional nonlinear orthogonal subspace. In this subspace, the topological neighborhoods of the feature submanifolds of refrigerant leakage and compressor anomalies do not intersect. This spatial orthogonal decoupling topological mechanism physically blocks the mutual interference of different fault features, ensuring that the input of subsequent models already possesses the basic conditions for fault separability.

[0063] At the adaptive reconstruction level, the local activation operators of the dynamically reconstructed local activation space evolution model maintain a prototype matrix internally. During forward propagation, the cosine similarity between the input feature vector and each prototype is used as a soft-gated weight for weighted aggregation, allowing the operator parameters to be dynamically reconstructed according to the local topological distribution of the input, rather than relying on fixed weights. This means that when the switching of operating conditions causes a shift in data distribution, the operator can adaptively adjust its internal representation space without changing the number of parameters, thereby suppressing parameter drift. The bidirectional backtracking gating unit monitors the entropy value of the hidden state distribution of adjacent layers in real time. Once an entropy surge is detected, the main forward path is immediately cut off and the process jumps to the memory trace decay branch. The memory trace decay branch performs incremental correction on the current layer weight matrix by exponentially decaying the historical state cache matrix. The correction amount is calculated by the outer product of the historical state cache matrix and the current hidden state. The contribution of recent states is greater than that of distant states, enabling the model to quickly recover diagnostic performance after a switching of operating conditions without full retraining. The physical constraint penalty term transforms the energy conservation equation and mass conservation equation of precision air conditioning into a residual penalty embedded loss function. The constraint model always follows the physical mechanism when extrapolating in the feature space, which further improves the generalization ability to unseen operating conditions.

[0064] At the operational condition perception level, the entropy drift adaptive adjustment function uses the Shannon entropy change rate of each operational condition parameter as the entropy drift index. Based on the interval to which the entropy drift index belongs, it triggers a tiered update strategy for local weights on the edge side. When the operational condition is stable, the weights remain unchanged to avoid unnecessary disturbances. When the operational condition changes drastically, the memory trace decay branch is activated to perform multi-step incremental corrections, forming an active suppression mechanism for parameter drift. These three levels work together to enable robust diagnosis of multi-source concurrent faults under cross-domain operational conditions without relying on full retraining.

[0065] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0066] The specific implementation method of step S01 is as follows.

[0067] The edge data acquisition unit synchronously collects vibration, refrigerant pressure, temperature field, and current signals from the precision air conditioning compressor. The sampling rate is determined based on the upper limit of the fault characteristic frequency band for each channel: no less than 10kHz for the vibration channel, no less than 1kHz for the pressure and temperature channels, and no less than 5kHz for the current channel. Unlocked ring queue capacity. Based on the measured peak sampling rate With reasoning delay The product is determined by the following formula:

[0068] ;

[0069] In the formula, This represents the capacity of the circular queue, in units of sampling points. The highest peak sampling rate in each channel, in units of , For inference delay, the unit is , This is a dimensionless redundancy margin coefficient, with an empirical value of 0.2. The dimension of the product of the above three factors is the sampling point, and... Consistent. Write pointer With read pointer All operations are performed through atomic comparison and swap operations, and the write operation satisfies the following conditions: When reading, satisfy This allows for thread-safe throughput without mutex locks. Packet loss rate Through statistical analysis of continuous pressure tests at different sampling frequencies, when Below the acceptable threshold Confirmation efficient, Based on the analysis of the cost of missed fault detection under the rated operating conditions of precision air conditioners, the output multi-channel timing data buffer is determined. ,in This represents the total number of sensor channels.

[0070] The specific implementation method of step S02 is as follows.

[0071] The multi-channel time-series data buffer is abstracted as a directed capacity network. ,in For a set of nodes, Let be a set of directed edges. This is a set of scalars representing edge capacities. Each sensor channel... Corresponding source node The sink node corresponding to the model input layer edge capacity By the Remaining bandwidth of the current channel buffer With occupancy The formula is determined together and expressed as follows:

[0072] ;

[0073] In the formula, For the first Node to the Edge capacity of a node, in units of , For the first Channel rated memory bandwidth, in units of , For the first Current buffer usage of the channel, in units of sampling points. As defined in step S01, This is a dimensionless bandwidth attenuation coefficient, with an empirical value of 1.0, and the exponential term... The attenuation factor is dimensionless, and the dimensions of both sides of the equation are 0. .exist The maximum flow is solved using the Ford-Fulkerson algorithm. The Ford-Fulkerson algorithm converges to the maximum flow value, corresponding to the minimum cut edge set, within a finite number of iterations by repeatedly searching for augmenting paths and updating the residual capacity on a directed capacity network. Directly indicates the current memory bandwidth bottleneck location; the algorithm is based on... The reading priority of high-frequency critical fault feature data stream is increased. The high-frequency critical fault feature data stream is determined as the channel where the fault feature energy is concentrated above 50Hz. The 50Hz threshold is determined by cross-validation experiments on the fault detection rate at different cutoff frequencies.

[0074] The specific implementation method of step S03 is as follows.

[0075] Manifold mapping is performed on multi-source heterogeneous data, projecting the original features onto a high-dimensional nonlinear orthogonal subspace using a Riemannian manifold framework. Let the original feature covariance matrix be... Two feature points and Geodesic distance between The definition is as follows:

[0076] ;

[0077] In the formula, The distance is a geodesic line, dimensionless. , Feature points , The corresponding covariance matrix, For matrix logarithm operations, Let be the Frobenius norm, and both sides of the equation be dimensionless scalars. The geodesic distance matrix for historical fault data. Perform singular value decomposition to obtain the orthogonal projection matrix. , For historical sample size, The projection subspace dimension is used to transform the original feature vector. Mapping to decoupled feature vectors The formula is expressed as follows:

[0078] ;

[0079] In the formula, This is the decoupled projected feature vector. This is the transpose of the orthogonal projection matrix. Since the original eigenvectors are constant, both sides of the equation have the same dimensions, inherited from the original eigenvectors. The dimensions of the submanifolds are such that the characteristic submanifolds of different failure modes have non-intersecting topological neighborhoods in the subspace, achieving stable decoupling of concurrent failures.

[0080] The specific implementation method of step S04 is as follows.

[0081] The decoupled normalized spacetime tensor Input a dynamic reconstruction local activation space evolution model. The time step, ranging from 64 to 512, was determined through experimental evaluation of the integrity of fault characteristics under different time windows. The streaming dynamic threshold normalization layer updates the mean of the current batch of data in real time. With variance Normalized output The formula is expressed as follows:

[0082] ;

[0083] In the formula, For the normalized spacetime tensor, This represents the average value of the current batch of data. The variance of the current batch of data. This is a numerically stable term, and the default value is... Both sides of the equation are dimensionless tensors.

[0084] Hidden layer A local activation operator, The value range is 4 to 12, the first... Each operator maintains a local prototype matrix. ,in This represents the number of prototypes, ranging from 8 to 64. For the feature dimension, the input feature vector is calculated during forward propagation. With the prototype vectors Cosine similarity, gate weight The formula is expressed as follows:

[0085] ;

[0086] Local activation output The formula is expressed as follows:

[0087] ;

[0088] In the formula, For the first Layer Dimensionless gating weights of a prototype For the summation index, and With consistent dimensions, the dimensions on both sides of the equation are the same as the prototype vector. Maintain consistency.

[0089] The bidirectional backtracking gating unit monitors the Shannon entropy of the hidden state distribution of adjacent layers in real time. ,when Exceeding the entropy surge threshold At this point, the gating unit cuts off the main forward path and jumps to the memory trace decay branch. The memory trace decay branch maintains the exponentially decaying historical state cache matrix. The updated formula is expressed as follows:

[0090] ;

[0091] In the formula, For the first Layer history state cache matrix, This is the attenuation coefficient, ranging from 0.85 to 0.99. The specific value is determined by applying different... The parameter drift of the model on a cross-seasonal working condition dataset was determined through multiple rounds of experiments. Let be the outer product matrix of the current hidden state vectors, and the dimensions of both sides of the equation are the same as . The outer product has consistent dimensions. Layer weight matrix. The incremental correction formula is expressed as follows:

[0092] ;

[0093] In the formula, For the first Layer weight matrix, The step size coefficient for dimensionless increment correction is empirically valued as follows: , It is an incremental correction quantity, and its dimensions are the same as those of the previous one. Consistent.

[0094] The loss function is the cross-entropy classification loss. Physical constraint penalty item It consists of two parts, and the formula is expressed as follows:

[0095] ;

[0096] In the formula, For the total loss, The cross-entropy classification loss is used, and all values ​​are dimensionless scalars. This is a dimensionless physical constraint weighting coefficient, with an empirical value of 0.1. The residual penalty term, used to transform the energy conservation equation and mass conservation equation of precision air conditioning, is specifically the weighted sum of squares of the system energy balance residual and the mass flow conservation residual. After normalization, it becomes a dimensionless scalar, with both sides of the equation being dimensionless. Adjacent hidden layers maintain consistent feature tensor dimensions through an adaptive step-size tensor concatenation operator. Output tensor from the previous layer The norm is determined, and the formula is expressed as follows:

[0097] ;

[0098] In the formula, For the first The layer splicing step size ranges from 1 to 4 and is a dimensionless integer. For the first Layer output tensors Norm, This is the step size adjustment coefficient, and its dimensions are... The reciprocal of the dimension ensures the product is dimensionless; the specific value depends on the training set. The quartiles of the distribution are determined. Model memory allocation is based on the bottleneck edge set of the maximum flow minimum cut model. capacity value As a ratio coefficient for memory allocation in each hidden layer, layers with larger bottleneck bandwidth are allocated more memory to support wider local prototype matrices, thus achieving linkage between memory resources and network flow scheduling parameters. The inter-layer neuron connection density is based on the variance of the gated weights output by the local activation operator of the previous layer. Adaptive pruning, when Below the variance threshold Reduce the number of connections to reduce redundant computation. Batch size was determined by quantile analysis of the variance distribution of gating weights on the training set. Based on the current real-time occupancy rate of the lock-free circular queue Dynamic adjustment: when occupancy is high, batch sizes are increased to accelerate consumption; when occupancy is low, batch sizes are decreased to reduce latency. This is a dimensionless occupancy rate.

[0099] The training dataset was established as follows: Refrigerant leakage, compressor bearing wear, compressor coil overheating, and concurrent combinations of these faults were simulated on a precision air conditioning test bench. At least 72 hours of multi-source sensor data were collected under different seasonal operating conditions, including full load in summer, low load in winter, and frequency conversion during the transition season. Professional technicians performed dual labeling of each data segment for fault type and operating condition type, forming a balanced sample set containing single faults, concurrent faults, and normal operating conditions. This set was then divided into training, validation, and test sets in a 7:1:2 ratio. Model training employed an adaptive weight decay optimizer to iteratively update the model parameters. The initial learning rate ranged from [value missing]. ~ The specific value is determined by the optimal convergence interval on the validation set loss curve through a learning rate warm-up experiment. After each training epoch, the hidden layer state entropy distribution is evaluated on the validation set and adjusted accordingly. Training is stopped and the optimal weights are saved when the diagnostic accuracy on the validation set does not improve within 10 consecutive rounds.

[0100] The specific implementation method of step S05 is as follows.

[0101] Entropy drift index The calculation formula is expressed as follows:

[0102] ;

[0103] In the formula, The dimensionless entropy drift exponent. This represents the total number of operating parameters, with a value of 4, corresponding to the four parameters: compressor speed, evaporating pressure, condensing pressure, and ambient temperature. For the first The dimensionless weights of each operating condition parameter satisfy the following: The principal component contribution rate was determined through experiments using historical operating condition switching data. and These are the current time window and the previous time window, respectively. The Shannon entropy of each operating parameter, in units of After weighted summation of the absolute value differences It is a dimensionless quantity. The rate of change of entropy of the operating condition parameters. For the first unit of time The rate of change of the Shannon entropy, a parameter of the operating condition, is The basic input for the calculation is expressed by the following formula:

[0104] ;

[0105] In the formula, For the first The rate of change of entropy of each operating parameter, in units of , The time interval between adjacent time windows, in units of , various and (Dimensionless) Multiplication and summation, overall For dimensionless quantities, the coefficient is... It already implies Normalization processing. Based on The local weights on the edge side of the interval are dynamically updated: when At that time, the local weights on the edge side of the model remain unchanged. Statistical analysis of fluctuation data under normal operating conditions on the test bench The 95th percentile is determined; when At that time, by step length (The range of values ​​is) ~ Perform single-step gradient descent incremental updates on the local weights at the edge; when At that time, the memory trace decay branch is activated, with a step size. (The range of values ​​is) ~ ,and Perform multi-step incremental corrections on all hidden layer weight matrices. By statistically analyzing the seasonal operating condition switching experiments on the test bench, we can determine the corresponding values ​​that cause a decrease in diagnostic accuracy exceeding acceptable losses. Quantiles are determined.

[0106] The specific implementation method of step S06 is as follows.

[0107] The multi-source fault diagnosis results output in step S04 are associated and stored with the equipment operation status record. The edge data acquisition unit outputs fault type identifiers and alarm commands according to the probability distribution of each fault type, thus completing the closed-loop diagnosis of multi-source faults in the precision air conditioner.

[0108] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0109] This embodiment uses a continuous operation scenario of a precision air conditioning system in a data center as an example to verify the diagnostic capability of the present invention for multi-source concurrent faults under cross-seasonal operating condition switching. The precision air conditioner has a rated cooling capacity of 50kW and is equipped with dual compressors. The main sensors include a compressor vibration acceleration sensor (range 0-50). Precision 0.01 The system includes refrigerant high and low pressure side pressure sensors (range 0–3 MPa, accuracy 0.005 MPa), evaporator and condenser temperature sensors (range -20–80℃, accuracy 0.1℃), and compressor drive current sensors (range 0–100A, accuracy 0.05A). The edge data acquisition unit is equipped with a quad-core processor and 512 MB of memory, enabling real-time acquisition of signals from these multiple sensors.

[0110] In step S01, the edge data acquisition unit uses four parallel acquisition threads to process the vibration (sampling rate 1000Hz), pressure (100Hz), temperature (50Hz), and current (500Hz) channels respectively. The lockless ring queue capacity is set based on twice the product of the peak sampling rate and the inference delay, with the vibration channel queue capacity set to 2000 sampling points, the current channel to 1000 sampling points, and the pressure and temperature channels each set to 200 sampling points. Each channel advances the read / write pointers using atomic operations. During a continuous 72-hour stress test, the waitless read / write pointer coordination mechanism ensured that the packet loss rate for each channel was below 0.05%, meeting the acceptable packet loss rate threshold.

[0111] In step S02, a directed capacity network is constructed using the four sensor channels as source nodes, the input layer of the dynamically reconstructed local activation space evolution model as the sink node, and the real-time occupancy and remaining bandwidth of each channel buffer as the edge capacity. After running the Ford-Fulkerson algorithm, the minimum cut edge set points to the vibration channel and the current channel, indicating that these are the current memory bandwidth bottlenecks. Power spectrum analysis results show that the characteristic energy of refrigerant leakage faults is concentrated in the range of 80–200 Hz, and the characteristic energy of compressor bearing wear faults is concentrated in the range of 120–500 Hz, both higher than the 50 Hz threshold. Both channels are identified as high-frequency critical fault characteristic data streams, and their memory read priority is higher than that of the pressure and temperature channels after scheduling. The scheduling priority allocation of each channel is shown in Table 1.

[0112] Table 1 Priority Allocation Table for Each Sensor Channel

[0113]

[0114] In step S03, the covariance matrix of each sensor channel within a time window of 256 sampling points is used as the feature representation. The geodesic distance matrix of the historical fault dataset (containing four categories: refrigerant leakage, compressor bearing wear, compressor coil overheating, and normal operating conditions, with at least 72 hours of data for each category) is calculated, and singular value decomposition is performed. The first 16 singular vectors are taken as orthogonal projection directions, and the original features are projected onto a 16-dimensional nonlinear orthogonal subspace. For example... Figure 2As shown, the topological neighborhoods of the refrigerant leakage feature submanifold and the compressor abnormality feature submanifold in the orthogonal subspace are clearly separated. Compared with the highly overlapping state in the original feature space, the decoupling effect is clearly visible. The spatial orthogonal decoupling topology mechanism effectively establishes the isolation boundary between fault types.

[0115] In step S04, the structural parameters of the dynamically reconstructed local activation space evolution model are shown in Table 2.

[0116] Table 2. Structural Parameters of the Dynamic Reconstruction Local Activation Spatial Evolution Model

[0117]

[0118] The model training dataset was divided into training, validation, and test sets in a 7:1:2 ratio, and trained using the AdamW optimizer with an initial learning rate of 100%. The learning rate was determined to be the median of the optimal convergence interval based on the validation set loss curve after a warm-up experiment. During the transitional season frequency conversion operation test, when the Shannon entropy of the evaporation and condensation pressures changed significantly within 5 minutes, the hidden layer state entropy increased from the normal value of approximately 1.2 nat to approximately 2.8 nat, exceeding the entropy surge threshold of 2.0 nat. The bidirectional backtracking gating unit then cut off the main forward path and activated the memory trace decay branch. After three incremental correction iterations, the hidden layer state entropy fell back to below 1.5 nat, and the model's diagnostic accuracy returned to the level before the operation switch, requiring no retraining throughout the process.

[0119] In step S05, the entropy drift index is calculated using the Shannon entropy of four operating parameters: compressor speed, evaporating pressure, condensing pressure, and ambient temperature. The weights of each parameter were determined experimentally based on the principal component contribution rate, and were 0.35, 0.28, 0.22, and 0.15, respectively. The first entropy drift threshold... Set to 0.18 (from the fluctuation data of the test bench under normal operating conditions) (obtained from the 95th percentile), second entropy drift threshold Set to 0.42 (the value corresponding to the decrease in diagnostic accuracy exceeding acceptable loss caused by seasonal operating condition switching experiments). (Obtained from the 90th percentile). For example... Figure 3 As shown, during the transition from summer to winter operating conditions, the entropy drift index... It gradually increases as the deviation in the distribution of operating parameters intensifies, when Enter When the interval is used, the length of the first step is used. Perform a single-step incremental update; when Rise to no less than At that time, with the second step length Activate multi-step incremental correction, local weights on the edge side follow The increase in the Shannon entropy of the operating condition parameter leads to a larger correction magnitude, suppressing parameter drift in the early stages of the operating condition switch, rather than triggering correction only after the diagnostic accuracy has significantly decreased. The physical significance of this mechanism is that the rate of change of the operating condition parameter Shannon entropy objectively reflects the migration speed of the data manifold. Using this as a trigger signal, intervention can be carried out in advance before the distribution offset has fully taken effect on the hidden layer weights, thus avoiding the lag problem of the fixed weight model detecting the decrease in accuracy only in the later stages of the operating condition switch.

[0120] In step S06, the diagnostic results, fault type probability distributions, and corresponding snapshots of operating parameters from the above steps are associated and stored. During the 48-hour continuous operation test in this embodiment, the system detected two refrigerant leakage events (both early minor leaks), one compressor bearing wear alarm, and one concurrent fault event involving both refrigerant leakage and bearing wear. All four events correctly generated fault type identifiers and promptly sent alarm commands, demonstrating the complete execution of the closed-loop diagnostic process. In the concurrent fault event, because the manifold mapping and spatial orthogonal decoupling topology mechanism have isolated the submanifolds of the two types of faults at the feature level, the dynamically reconstructed local activation space evolution model can simultaneously output high-confidence probabilities for both types of faults, avoiding the phenomenon of missed detections due to mutual masking of concurrent faults in traditional schemes. This result indicates that when the input features are topologically separable, the classification layer only needs to independently estimate the probability distribution of each submanifold, without forcibly distinguishing them in the aliased feature space, thus effectively ensuring the diagnostic confidence of concurrent faults.

[0121] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for multi-source fault diagnosis of precision air conditioning based on edge data acquisition devices, characterized in that, Includes the following steps: The edge data acquisition unit collects compressor vibration, refrigerant pressure, temperature field and current signals of precision air conditioners. Through the lockless ring queue and waitless read and write pointer collaborative mechanism, it completes high-concurrency and robust throughput of multi-source high-frequency streaming data and outputs multi-channel time-series data buffer. The multi-channel time-series data buffer is abstracted into a network flow topology graph with limited capacity. The memory read bandwidth and computation priority of each sensor data stream are dynamically allocated by solving the maximum flow minimum cut model. Manifold mapping is performed on the multi-source heterogeneous data that has been scheduled, projecting the original features onto a high-dimensional nonlinear orthogonal subspace, establishing a spatial orthogonal decoupling topology mechanism, and achieving precise decoupling and isolation of multi-source concurrent fault features; The decoupled standardized spatiotemporal tensor is input into the dynamic reconstruction of the local activation space evolution model. The model dynamically reconstructs the hidden layer operator parameters based on the local topological distribution of the input data, and the distribution offset is detected by a bidirectional backtracking gating mechanism. The memory trace decay branch is called to perform incremental correction on the layer weight matrix and output multi-source fault diagnosis results. The entropy drift adaptive adjustment function is calculated based on the entropy change rate under the operating conditions to obtain the entropy drift index. The edge local weights of the dynamically reconstructed local activation space evolution model are then dynamically updated based on the interval to which the entropy drift index belongs. The multi-source fault diagnosis results are associated with and stored with the equipment operating status records, and the fault type identifier and alarm command are output through the edge data acquisition device.

2. The method for multi-source fault diagnosis of precision air conditioning based on edge data acquisition unit according to claim 1, characterized in that, The lock-free circular queue refers to replacing the traditional locked queue with a fixed-capacity circular buffer. Producer and consumer threads each hold independent write pointers and read pointers, and there is no mutual exclusion waiting between the two pointers. The queue capacity is determined by the product of the measured peak sampling rate and the inference latency, and the packet loss rate is adjusted to below an acceptable threshold through iterative stress testing.

3. The method for multi-source fault diagnosis of precision air conditioning based on edge data acquisition unit according to claim 2, characterized in that, The aforementioned wait-free read-write pointer coordination mechanism uses atomic operations to advance both the write pointer and the read pointer, allowing any thread to complete read and write operations directly at any time without waiting for other threads to release resources.

4. The method for multi-source fault diagnosis of precision air conditioning based on edge data acquisition unit according to claim 3, characterized in that, The capacity-constrained network flow topology graph uses each sensor channel as the source node, the input layer of the dynamic reconstruction local activation space evolution model as the sink node, and the current occupancy and remaining bandwidth of each channel's data buffer as the edge capacity to form a directed capacity network. The maximum flow minimum cut model is solved by the Ford-Fulkerson algorithm, and the edge set corresponding to the minimum cut is the bottleneck edge set of the memory bandwidth at the current moment. The determination of high-frequency key fault feature data flow is based on the frequency band where the fault feature energy is concentrated.

5. The method for multi-source fault diagnosis of precision air conditioning based on edge data acquisition unit according to claim 4, characterized in that, The manifold mapping utilizes a nonlinear dimensionality reduction operator based on the Riemannian manifold framework, using the geodesic distance of the covariance matrix as a feature similarity measure. By performing singular value decomposition on the geodesic distance matrix of historical fault data, the orthogonal projection direction is determined, and multiple fault features are mapped to a high-dimensional nonlinear orthogonal subspace.

6. The method for multi-source fault diagnosis of precision air conditioning based on edge data acquisition unit according to claim 5, characterized in that, The spatial orthogonal decoupling topology mechanism establishes isolation boundaries based on the topological connectivity of the characteristic distribution of each fault type, so that the topological neighborhoods of the refrigerant leakage characteristic submanifold and the compressor abnormal characteristic submanifold in the subspace do not intersect.

7. The method for multi-source fault diagnosis of precision air conditioning based on edge data acquisition unit according to claim 6, characterized in that, The dynamically reconstructed local activation space evolution model has an input layer receiving dimension of... The standardized spacetime tensor, in which Number of sensor channels The time step is specified; the input layer is followed by a streaming dynamic threshold normalization layer.

8. The method for multi-source fault diagnosis of precision air conditioning based on edge data acquisition unit according to claim 7, characterized in that, The internal maintenance dimension of the local activation operator is: The local prototype matrix is ​​calculated during forward propagation, and the cosine similarity between the input feature vector and each prototype is calculated. The similarity is used as a soft gating weight to perform weighted aggregation of the prototype matrix, so as to realize the dynamic reconstruction of the operator parameters with the local topological distribution of the input. The spatial self-organizing evolution mechanism of the prototype matrix moves the prototype towards the local density center of the input data through competitive learning during the training phase.

9. The method for multi-source fault diagnosis of precision air conditioning based on edge data acquisition unit according to claim 8, characterized in that, The bidirectional backtracking gating unit monitors the entropy value of the hidden state distribution of adjacent layers in real time. When the entropy of the hidden state exceeds the entropy surge threshold, the main forward path is cut off and the process jumps to the memory trace decay branch. The memory trace decay branch maintains an exponentially decaying historical state cache matrix. The historical state cache matrix is ​​used to perform incremental correction on the current layer weight matrix. The correction amount is calculated by the outer product of the historical state cache matrix and the current hidden layer state.

10. The method for multi-source fault diagnosis of precision air conditioning based on edge data acquisition unit according to claim 9, characterized in that, The network loss function of the dynamically reconstructed local activation spatial evolution model consists of two parts: cross-entropy classification loss and physical constraint penalty term. The physical constraint penalty term transforms the energy conservation equation and mass conservation equation of the precision air conditioning system into residual penalty. Adjacent hidden layers maintain consistent feature tensor dimensions through an adaptive step-size tensor splicing operator. The splicing operator is based on the output tensor of the current layer. The norm adaptively adjusts the step size.

Images